DATA PIPELINE AND DEEP LEARNING SYSTEM FOR AUTONOMOUS

DRIVING

BACKGROUND OF THE INVENTION

[0001] Deep learning systems used to implement autonomous driving typically rely on captured sensor data as input. In traditional learning systems, the captured sensor data is made compatible with a deep learning system by converting the captured data from a sensor format to a format compatible with the initial input layer of the learning system. This conversion may include compression and down-sampling that can reduce the signal fidelity of the original sensor data. Moreover, changing sensors may require a new conversion process. Therefore, there exists a need for a customized data pipeline that can maximize the signal information from the captured sensor data and provide a higher level of signal information to the deep learning network for deep learning analysis.

BRIEF DESCRIPTION OF THE DRAWINGS

[0002] Various embodiments of the invention are disclosed in the following detailed description and the accompanying drawings.

[0003] Figure 1 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using a deep learning pipeline.

[0004] Figure 2 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using a deep learning pipeline.

[0005] Figure 3 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using component data.

[0006] Figure 4 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using high-pass and low-pass component data.

[0007] Figure 5 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using high-pass, band-pass, and low-pass component data.

[0008] Figure 6 is a block diagram illustrating an embodiment of a deep learning system for

autonomous driving.

DETAILED DESCRIPTION

[0009] The invention can be implemented in numerous ways, including as a process; an apparatus; a system; a composition of matter; a computer program product embodied on a computer readable storage medium; and/or a processor, such as a processor configured to execute instructions stored on and/or provided by a memory coupled to the processor. In this specification, these implementations, or any other form that the invention may take, may be referred to as techniques.

In general, the order of the steps of disclosed processes may be altered within the scope of the invention. Unless stated otherwise, a component such as a processor or a memory described as being configured to perform a task may be implemented as a general component that is temporarily configured to perform the task at a given time or a specific component that is manufactured to perform the task. As used herein, the term‘processor’ refers to one or more devices, circuits, and/or processing cores configured to process data, such as computer program instructions.

[0010] A detailed description of one or more embodiments of the invention is provided below along with accompanying figures that illustrate the principles of the invention. The invention is described in connection with such embodiments, but the invention is not limited to any embodiment. The scope of the invention is limited only by the claims and the invention encompasses numerous alternatives, modifications and equivalents. Numerous specific details are set forth in the following description in order to provide a thorough understanding of the invention. These details are provided for the purpose of example and the invention may be practiced according to the claims without some or all of these specific details. For the purpose of clarity, technical material that is known in the technical fields related to the invention has not been described in detail so that the invention is not unnecessarily obscured.

[0011] A data pipeline that extracts and provides sensor data as separate components to a deep learning network for autonomous driving is disclosed. In some embodiments, autonomous driving is implemented using a deep learning network and input data received from sensors. For example, sensors affixed to a vehicle provide real-time sensor data, such as vision, radar, and ultrasonic data, of the vehicle's surrounding environment to a neural network for determining vehicle control responses. In some embodiments, the network is implemented using multiple layers. The sensor data is extracted into two or more different data components based on the signal information of the data. For example, feature and/or edge data may be extracted separate from

global data such as global illumination data into different data components. The different data components retain the targeted relevant data, for example, data that will eventually be used to identify edges and other features by a deep learning network. In some embodiments, the different data components function as containers that store data highly relevant for identifying certain targeted features but do not themselves identify or detect the features. The different data components extract data to ensure accurate feature detection at an appropriate stage of a machine learning network. In some embodiments, the different data components may then be pre-processed to enhance the particular signal information that they contain. The data components can be compressed and/or down-sampled to increase resource and computational efficiency. The different data components are then provided to the deep learning system at different layers of the system.

The deep learning network is able to accurately identify and detect the features associated with the targeted data (e.g., edges, objects, etc.) of the data component using the signal information retained during extraction as input. For example, feature and edge data is provided to the first layer of the network and global data to a later layer of the network. By extracting different data components that each retain their respective targeted signal information, the network more efficiently processes the sensor data. Instead of receiving the sensor data as an initial input to the network, the network is provided with the most useful information at the most appropriate layer of the network. In some embodiments, a more complete version of the captured sensor data is analyzed by the network since the different data components can fully utilize the image resolution of their respective components for their intended purposes. For example, input for features and edges can utilize the entire resolution, bit range, and bit depth for feature and edge data whereas input for global illumination can utilize the entire resolution, bit range, and bit depth for global illumination data.

[0012] In some embodiments, an image captured using a sensor on a vehicle is received.

For example, an image is captured from a high dynamic range forward-facing camera. As another example, ultrasonic data is captured from a side-facing ultrasonic sensor ln some embodiments, the received image is decomposed into a plurality of component images. For example, feature data is extracted from a captured high dynamic range image. As another example, global illumination data is extracted from the captured high dynamic range image. As another example, the image may be decomposed using high-pass, low-pass, and/or band-pass filters ln some embodiments, each component image of the plurality of component images is provided as a different input to a different layer of a plurality of layers of an artificial neural network to determine a result. For example, an artificial neural network such as a convolutional neural network includes multiple layers for processing input data. The different component images decomposed from the captured image are presented as input to different layers of the neural network. For example, feature data is presented as input to the first layer of the network and global data is presented as input to a later layer (e.g., the third layer) of the network. In some embodiments, the result of the artificial neural network is used to at least in part autonomously operate the vehicle. For example, the result of deep learning analysis using the artificial neural network is used to control the steering, breaking, lighting, and/or warning systems of the vehicle. In some embodiments, the result is used to autonomously match the vehicle's speed to traffic conditions, steer the vehicle to follow a navigational path, avoid collisions when an object is detected, summon the vehicle to a desired location, and warn the user of potential collisions, among other autonomous driving applications.

[0013] In some embodiments, a vehicle is affixed with multiple sensors for capturing data.

For example, in some embodiments, eight surround cameras are affixed to a vehicle and provide 360 degrees of visibility around the vehicle with a range of up to 250 meters. In some embodiments, camera sensors include a wide forward camera, a narrow forward camera, a rear view camera, forward looking side cameras, and/or rearward looking side cameras. In some embodiments, ultrasonic and radar sensors are used to capture surrounding details. For example, twelve ultrasonic sensors may be affixed to the vehicle to detect both hard and soft objects. In some embodiments, a forward-facing radar is utilized to capture data of the surrounding environment. In various embodiments, radar sensors are able to capture surrounding detail despite heavy rain, fog, dust, and other vehicles. The various sensors are used to capture the environment surrounding the vehicle and the captured image is provided for deep learning analysis.

[0014] Using data captured from sensors and analyzed using the disclosed deep learning system, a machine learning result is determined for autonomous driving. In various embodiments, the machine learning result is provided to a vehicle control module for implementing autonomous driving features. For example, a vehicle control module can be used to control the steering, braking, warning systems, and/or lighting of the vehicle. In some embodiments, the vehicle is controlled to navigate roads, match the speed of the vehicle to traffic conditions, keep the vehicle within a lane, automatically change lanes without requiring driver input, transition the vehicle from one freeway to another, exit the freeway when approaching a destination, self-park the vehicle, and summon the vehicle to and from a parking spot, among other autonomous driving applications. In some embodiments, the autonomous driving features include identifying opportunities to move the vehicle into a faster lane when behind slower traffic. In some embodiments, the machine learning result is used determine when autonomous driving without driver interaction is appropriate and when it should be disabled. In various embodiments, the machine learning result is used to assist a driver in driving the vehicle.

[0015] In some embodiments, the machine learning result is used to implement a self parking mode where the vehicle will automatically search for a parking spot and park the vehicle.

In some embodiments, the machine learning result is used to navigate the vehicle using a destination from a user's calendar. In various embodiments, the machine learning result is used to implement autonomous driving safety features such as collision avoidance and automatic emergency braking. For example, in some embodiments, the deep learning system detects objects that may impact with the vehicle and the vehicle control module applies the brakes accordingly. In some embodiments, the vehicle control module uses the deep learning analysis to implement a side, front, and/or rear collision warning that warns the user of the vehicle of potential collisions with obstacles alongside, in front, or behind the vehicle. In various embodiments, the vehicle control module can activate warning systems such as collision alerts, audio alerts, visual alerts, and/or physical alerts (such as vibration alerts), among others, to inform the user of an emergency situation or when the driver's attention is necessary. In some embodiments, the vehicle control module can initiate a communication response such as an emergency response call, a text message, a network update, and/or another communication response as appropriate, for example, to inform another party of an emergency situation. In some embodiments, the vehicle control module can adjust the lighting including the high/low beams, brake lights, interior light, emergency lights, etc. based on the deep learning analysis results. In some embodiments, the vehicle control module can further adjust the audio in or around the vehicle including using the horn, modifying the audio (e.g., music, phone calls, etc.) playing from the vehicle's sound system, adjusting the volume of the sound system, playing audio alerts, enabling a microphone, etc. based on deep learning analysis results.

[0016] Figure 1 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using a deep learning pipeline. For example, the process of Figure 1 may be utilized to implement autonomous driving features for self-driving and driver-assisted automobiles to improve safety and to reduce the risk of accidents. In some embodiments, the process of Figure 1 pre-processes data captured by sensors for deep learning analysis. By pre processing the sensor data, the data provided for deep learning analysis is enhanced and results in a more accurate result for controlling a vehicle. In some embodiments, the pre-processing addresses data mismatches between data captured by sensors and data expected by a neural network for deep learning.

[0017] At 101, sensor data is received. For example, sensor data is captured by one or more sensors affixed to a vehicle. In some embodiments, the sensors are affixed to the environment

and/or other vehicles and data is received remotely. In various embodiments, the sensor data is image data, such as RGB or YUV channels of an image. In some embodiments, the sensor data is captured using a high dynamic range camera. In some embodiments, the sensor data is radar, LiDAR, and/or ultrasonic data. In various embodiments, LiDAR data is data captured using laser light and may includes techniques referred to as Light Detection And Ranging as well as Laser Imaging, Detection and Ranging. In various embodiments, the bit depth of the sensor data exceeds the bit depth of the neural network for deep learning analysis.

[0018] At 103, data pre-processing is performed on the sensor data. In some embodiments, one or more pre-processing passes may be performed on the sensor data. For example, the data may be first pre-processed to remove noise, to correct for alignment issues and/or blurring, etc. In some embodiments, two or more different filtering passes are performed on the data. For example, a high-pass filter may be performed on the data and a low-pass filter may be performed on the data. In some embodiments, one or more band pass filters may be performed. For example, one or more band passes may be performed on the data in addition to a high-pass and a low-pass. In various embodiments, the sensor data is separated into two or more data sets such as a high-pass data set and a low-pass data set. In some embodiments, one or more band pass data sets are also created.

In various embodiments, the different data sets are different components of the sensor data.

[0019] In some embodiments, the different components created by pre-processing the data include a feature and/or edge component and a global data component. In various embodiments, the feature and/or edge component is created by performing a high-pass or band-pass filter on the sensor data and the global data component is created by performing a low-pass or band-pass filter on the sensor data. In some embodiments, one or more different filter techniques may be used to extract feature/edge data and/or global data.

[0020] In various embodiments, one or more components of the sensor data are processed.

For example, a high-pass component may be processed by removing noise from and/or enhancing local contrast for the image data. In some embodiments, the low-pass component is compressed and/or down-sampled. In various embodiments, different components are compressed and/or down-sampled. For example, components may be compressed, resized, and/or down-sampled as appropriate to adjust the size and/or resolution of the data for inputting the data to a layer of a machine learning model. In some embodiments, the bit depth of the sensor data is adjusted. For example, a data channel of a camera capturing data at 20-bits or another appropriate bit depth is compressed or quantized to 8-bits to prepare the channel for an 8-bit machine learning model. In some embodiments, one or more sensors capture data at a bit depth of l2-bits, 16-bits, 20-bits, 32- bits, or another appropriate bit depth that is larger than the bit depth used by the deep learning network.

[0021] In various embodiments, the pre-processing performed at 103 is performed by an image pre-processor. In some embodiments, the image pre-processor is a graphics processing unit (GPU), a central processing unit (CPU), an artificial intelligence (AI) processor, an image signal processor, a tone-mapper processor, or other similar hardware processor. In various embodiments, different image pre-processors are used to extract and/or pre-process different data components in parallel.

[0022] At 105, deep learning analysis is performed. For example, deep learning analysis is performed using a machine learning model such as an artificial neural network. In various embodiments, the deep learning analysis receives the processed sensor data for 103 as input. In some embodiments, the processed sensor data is received at 105 as multiple different components, such as a high-pass data component and a low-pass data component. In some embodiments, the different data components are received as inputs to different layers of the machine learning model. For example, a neural network receives a high-pass component as an initial input to the first layer of the network and a low-pass component as input to a subsequent layer of the network.

[0023] At 107, the results of the deep learning analysis are provided for vehicle control.

For example, the results may be provided to a vehicle control module to adjust the speed and/or steering of the vehicle. In various embodiments, the results are provided to implement autonomous driving functionality. For example, the results may indicate an object that should be avoided by steering the vehicle. As another example, the results may indicate a merging car that should be avoided by braking and changing the vehicle's positioning in the lane.

[0024] Figure 2 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using a deep learning pipeline. For example, the process of Figure 2 may be utilized to pre-process sensor data, extract image components from the sensor data, pre-process the extracted image components, and then provide the components for deep learning analysis. The results of deep learning analysis may be used to implement autonomous driving to improve safety and to reduce the risk of accidents. In some embodiments, the process of Figure 2 is used to perform the process of Figure 1. In some embodiments, step 201 is performed at 101 of Figure 1; steps 203, 205, 207, and/or 209 are performed at 103 of Figure 1; and/or step 211 is performed at 105 and/or 107 of Figure 1. By processing the extracted components of the sensor data, the processed data provided to a machine learning model is enhanced to achieve superior results from deep learning analysis rather than using otherwise non-enhanced data. In the example shown, the results of the deep learning analysis are used for vehicle control.

[0025] At 201, sensor data is received. In various embodiments, the sensor data is image data captured from a sensor such as a high dynamic range camera. In some embodiments, the sensor data is captured from one or more different sensors. In some embodiments, the image data is captured using a l2-bit or higher bit depth to increase the fidelity of the data.

[0026] At 203, the data is pre-processed. In some embodiments, the data is pre-processed using an image pre-processor such as an image signal processor, a graphics processing unit (GPU), a tone-mapper processor, a central processing unit (CPU), an artificial intelligence (AI) processor, or other similar hardware processor. In various embodiments, linearization, demosaicing, and/or another processing techniques may be performed on the captured sensor data. In various embodiments, pre-processing is performed on the high-resolution sensor data to enhance the fidelity of the captured data and/or to reduce the introduction of errors by subsequent steps. In some embodiments, the pre-processing step is optional.

[0027] At 205, one or more image components are extracted. In some embodiments, two image components are extracted. For example, a feature/edge data component of the sensor data is extracted and a global data component of the sensor data is extracted. In some embodiments, a high-pass component and a low-pass component of the sensor data are extracted. In some embodiments, one or more additional band-pass components are extracted from the sensor data. In various embodiments, high-pass, low-pass, and/or band-pass filters are used to extract different components of the sensor data. In some embodiments, the image components are extracted using a tone mapper. In some embodiments, the global data and/or low-pass component data is extracted by down-sampling the sensor data using a binning or similar technique. In various embodiments, the extraction retains and saves the targeted signal information as an image data component but does not actually detect or identify the features related to the targeted information. For example, the extraction of an image component corresponding to edge data results in an image component with targeted signal information for accurately identifying edges but the extraction performed at 205 does not detect the existence of edges in the sensor data.

[0028] In some embodiments, the image data component extracted for a first layer of a machine learning network is extracted using a process that preserves the response of the first layer of the deep learning analysis. For example, the relevant signal information for the first layer is preserved such that the result of the analysis performed on the image component after the analysis of the first layer is similar to the analysis performed on the corresponding sensor data prior to extraction into image components. In various embodiments, the results are preserved for filters as small as a 5x5 matrix filter.

[0029] In some embodiments, an extracted data component is created by combining multiple channels of the captured image into one or more channels. For example, red, green, and blue channels may be averaged to create a new channel for a data component. In various embodiments, an extracted data component may be constructed from one or more different channels of the source capture data and/or one or more different captured images of different sensors. For example, data from multiple sensors may be combined into a single data component.

[0030] In some embodiments, an image pre-processor such as the pre-processor of step 203 is used to extract the different components. In some embodiments, an image signal processor may be used to extract the different components. In various embodiments, a graphics processing unit (GPU) may be used to extract the different components. In some embodiments, a different pre processor is used to extract different components so that multiple components can be extracted in parallel. For example, an image signal processor may be used to extract a high-pass component and a GPU may be used to extract a low-pass component. As another example, an image signal processor may be used to extract a low-pass component and a GPU may be used to extract a high-pass component. In some embodiments, a tone-mapper processor is used to extract an image component (such as a high-pass component) and a GPU is used to extract a separate image component (such as a low-pass component) in parallel. In some embodiments, the tone-mapper is part of an image signal processor. In some embodiments, multiple instances of similar preprocessors exist to perform extractions in parallel.

[0031] At 207, component pre-processing is performed. In some embodiments, an image pre-processor such as the pre-processor of step 203 and/or 205 is used to pre-process one or more components. In some embodiments, a different pre-processor is used to pre-process different components so that the pre-processing can be performed on the different components in parallel. For example, an image signal processor may be used to process a high-pass component and a graphics processing unit (GPU) may be used to process a low-pass component. In some embodiments, a tone-mapper processor is used to process one image component and a GPU is used to process a separate image component in parallel. In some embodiments, multiple instances of similar pre-processors exist for processing different components in parallel.

[0032] In various embodiments, the pre-processing includes down-sampling and/or

compressing the image component data. In some embodiments, the pre-processing includes removing noise from the component data. In some embodiments, the pre-processing includes compressing or quantizing the captured data from 20-bit down to 8-bit data fields. In some embodiments, the pre-processing includes converting the size of the image component to a lower resolution. For example, an image component may be half, a quarter, an eighth, a sixteenth, one thirty-second, one sixty-fourth, or another appropriate scaling of the original sensor image size. In various embodiments, an image component is reduced to a size appropriate to the input layer of the machine learning model.

[0033] At 209, components are provided to the appropriate network layer of the deep learning network. For example, different components may be provided to different layers of the network. In some embodiments, the network is a neural network with multiple layers. For example, the first layer of a neural network receives as input high-pass component data. One of the subsequent network layers receives as input low-pass component data corresponding to global illumination data. In various embodiments, the different components extracted at 205 and pre-processed at 207 are received at different layers of the neural network. As another example, a feature and/or edge data component is provided as input to the first layer of a deep learning network such as an artificial neural network. A global data component is provided to a subsequent layer and can be provided as a compressed and/or down-sampled version of the data since the global data does not require as much precision as feature and/or edge component data. In various embodiments, global data is more easily compressed without losing information and can be provided at a later layer of the network.

[0034] In some embodiments, the machine learning model is made up of multiple sequential layers where the one or more subsequent layers receive input data that has a size property that is smaller in size than a previous layer. For example, the first layer to a network may receive an image size similar to the capture image size. Subsequent layers may receive input data that is a half or a quarter of the capture image size. The reduction in input data size reduces the computation of subsequent layers and improves the efficiency of the deep learning analysis. By providing the sensor input data as different components and at different layers, computational efficiency is increased. Earlier layers of the network require increased computation in particular because the amount of data and the data size is larger than subsequent layers. Subsequent layers may be more efficient to compute since the input data has been compressed by previous layers of the network and/or the pre-processing at 207.

[0035] At 211, results of the deep learning analysis are provided for vehicle control. For

example, machine learning results using the processed image components may be utilized to control a vehicle's movement. In some embodiments, the results correspond to vehicle control actions. For example, results may correspond to the speed and steering of the vehicle. In some embodiments, the results are received by a vehicle control module used to help maneuver the vehicle. In some embodiments, the results are utilized to improve the safety of the vehicle. In various embodiments, the results provided at 211 are determined by performing a deep learning analysis on the components provided at 209.

[0036] Figure 3 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using component data. In the example shown, the process of Figure 3 is used to extract feature and edge data from sensor data separate from global data. The two data sets are then fed into a deep learning network at different stages to infer vehicle control results. By separating the two components and providing them at different stages, the initial layers of the network can dedicate computational resources to initial edges and feature detection. In some embodiments, the initial stages dedicate resources to the initial identification of the objects such as roads, lane markers, obstacles, vehicles, pedestrians, traffic signs, etc. Subsequent layers can utilize the global data in a more computational efficient manner since the global data is less resource intensive. Since machine learning can be computational and data intensive, a data pipeline utilizing different image components at different stages is utilized to increase the efficiency of the deep learning computation and to reduce data resource requirements needed for the analysis. In some embodiments, the process of Figure 3 is used to perform the process of Figure 1 and/or Figure 2. In some embodiments, step 301 is performed at 101 of Figure 1 and/or at 201 of Figure 2; step 303 is performed at 103 of Figure 1 and/or at 203 of Figure 2; steps 311 and/or 321 are performed at 103 of Figure 1 and/or at 205 of Figure 2; steps 313, 323, and/or 325 are performed at 103 of Figure 1 and/or at 207 and 209 of Figure 2; steps 315 and/or 335 are performed at 105 of Figure 1 and/or at 211 of Figure 2; and/or step 337 is performed at 107 of Figure 1 and/or at 211 of Figure 2.

[0037] At 301, sensor data is received. In various embodiments, the sensor data is data captured by one or more sensors of the vehicle. In some embodiments, the sensor data is received as described with respect to step 101 of Figure 1 and/or step 201 of Figure 2.

[0038] At 303, data pre-processing is performed. For example, the sensor data is enhanced by pre-processing the data. In some embodiments, the data is cleaned up, for example, by performing a de-noising, alignment, or other appropriate filter. In various embodiments, the data is pre-processed as described with respect to step 103 of Figure 1 and/or step 203 of Figure 2. In the example shown, processing continues to steps 311 and 321. In some embodiments, processing at

311 and 321 are run in parallel to extract and process different components of the sensor data. In some embodiments, each branch of processing (e.g., the branch starting at 311 and the branch starting at 321) is run sequentially or pipelined. For example, processing is performed starting with step 311 to prepare data for the initial layers of a network. In some embodiments, the preprocessing step is optional.

[0039] At 311, feature and/or edge data is extracted from the sensor data. For example, feature data and/or edge data is extracted from the captured sensor data into a component data. In some embodiments, the component data retains the relevant signal information from the sensor data for identifying features and/or edges. In various embodiments, the extraction process preserves the signal information critical for identifying and detecting features and/or edges and does not actually identify or detect the features or edges from the sensor data. In various embodiments, the features and/or edges are detected during one or more analysis steps at 315 and/or 335. In some embodiments, the extracted feature and/or edge data has the same bit depth as the original captured data. In some embodiments, the extracted data is feature data, edge data, or a combination of feature and edge data. In some embodiments, a high-pass filter is used to extract feature and/or edge data from the sensor data. In various embodiments, a tone-mapper processor is calibrated to extract feature and/or edge data from the sensor data.

[0040] At 313, pre-processing is performed on the feature and/or edge data. For example, a de-noising filter may be applied to the data to improve the signal quality. As another example, different pre-processing techniques such as local contrast enhancement, gain adjustment, thresholding, noise filtering, etc. may be applied to enhance the feature and edge data prior to deep learning analysis. In various embodiments, the pre-processing is customized to enhance the feature and edge properties of the data rather than applying a more generic pre-processing technique to the sensor data as a whole. In some embodiments, the pre-processing includes performing a compression and/or down-sampling on the extracted data. In some embodiments, the preprocessing step at 313 is optional.

[0041] At 315, an initial analysis is performed using the feature and/or edge data. In some embodiments, the initial analysis is a deep learning analysis using a machine learning model such as a neural network. In various embodiments, the initial analysis receives the feature and edge data as input to the first layer of the network. In some embodiments, the initial layer of the network prioritizes the detection of features and/or edges in the captured image. In various embodiments, the deep learning analysis is performed using an artificial neural network such as a convolutional neural network. In some embodiments, the analysis is run on an artificial intelligence (AI)

processor.

[0042] At 321, global data is extracted from the sensor data. For example, global data is extracted from the captured sensor data into a component data. In some embodiments, the global data corresponds to global illumination data. In some embodiments, the extracted global data has the same bit depth as the original captured data. In some embodiments, a low -pass filter is used to extract global data from the sensor data. In various embodiments, a tone-mapper processor is calibrated to extract global data from the sensor data. Other techniques, such as binning, resampling, and down-sampling may also be used to extract global data. In various embodiments, the extraction process retains data likely to be globally relevant and does not identify and detect the global features from the sensor data. In various embodiments, the global features are detected by the analysis performed at 335.

[0043] At 323, pre-processing is performed on the global data. For example, a de-noising filter may be applied to the data to improve the signal quality. As another example, different pre processing techniques such as local contrast enhancement, gain adjustment, thresholding, noise filtering, etc. may be applied to enhance the global data prior to deep learning analysis. In various embodiments, the pre-processing is customized to enhance the properties of the global data rather than applying a more generic pre-processing technique to the sensor data as a whole. In some embodiments, the pre-processing of the global data includes compressing the data. In some embodiments, the pre-processing step at 323 is optional.

[0044] At 325, the global data is down-sampled. For example, the resolution of the global data is reduced. In some embodiments, the global data is reduced in size to improve the computational efficiency of analyzing the data and to configure the global data as input to a later layer of the deep learning network. In some embodiments, the global data is down-sampled by binning, resampling, or another appropriate technique. In some embodiments, the down-sampling is performed using a graphical processing unit (GPU) or an image signal processor. In various embodiments, down-sampling is appropriate for global data since the global data does not have the same resolution requirements as feature and/or edge data. In some embodiments, the down-sampling performed at 325 is performed when the global data is extracted at 321.

[0045] At 335, additional deep learning analysis is performed using results of the deep learning analysis on the feature and/or edge data and the global data as input. In various embodiments, the deep learning analysis receives as input the global data at a later layer of the deep learning network. In various embodiments, the expected input data size at the layer receiving the global data is smaller than the expected input data size of the initial input layer. For example, the input size for the global data input layer may be a half or a quarter of the input size for the initial layer of the deep learning network. In some embodiments, the later layers of the network utilize global data to enhance the results of the initial layers. In various embodiments, the deep learning analysis is performed and a vehicle control result is determined. For example, a vehicle control result is determined using a convolutional neural network. In some embodiments, the analysis is run on an artificial intelligence (AI) processor.

[0046] At 337, the results of deep learning analysis are provided for vehicle control. For example, machine learning results using the extracted and processed image components are utilized to control a vehicle's movement. In some embodiments, the results correspond to vehicle control actions. In some embodiments, the results are provided as described with respect to step 107 of Figure 1 and/or step 211 of Figure 2.

[0047] Figure 4 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using high-pass and low-pass component data. In the example shown, the process of Figure 4 is used to extract two data components from sensor data and to provide the components to different layers of a deep learning network such as an artificial neural network. The two components are extracted using a high-pass and low-pass filter. In various embodiments, the results are used to implement autonomous driving with improved precision, safety, and/or comfort results. In some embodiments, the process of Figure 4 is used to perform the process of Figures 1, 2, and/or 3. In some embodiments, step 401 is performed at 103 of Figure 1, at 203 of Figure 2, and/or at 303 of Figure 3; step 403 is performed at 103 of Figure 1, at 205 of Figure 2, and/or at 311 of Figure 3; step 413 is performed at 103 of Figure 1, at 205 of Figure 2, and/or at 321 of Figure 3; step 405 is performed at 103 of Figure 1, at 207 and 209 of Figure 2, and/or at 313 of Figure 3; steps 415 and 417 are performed at 103 of Figure 1, at 207 and 209 of Figure 2, and/or at 323 and 325 of Figure 3; step 407 is performed at 105 of Figure 1, at 211 of Figure 2, and/or at 315 of Figure 3; and/or step 421 is performed at 105 of Figure 1, at 211 of Figure 2, and/or at 335 of Figure 3.

[0048] At 401, data is pre-processed. In some embodiments, the data is the sensor data captured from one or more sensors such as high dynamic range camera, radar, ultrasonic, and/or LiDAR sensors. In various embodiments, the data is pre-processed as described with respect to 103 of Figure 1, 203 of Figure 2, and/or 303 of Figure 3. Once the data is pre-processed, processing continues to 403 and 413. In some embodiments, steps 403 and 413 are run in parallel.

[0049] At 403, a high-pass filter is performed on the data. For example, a high-pass filter is performed on the captured sensor data to extract high-pass component data. In some embodiments, the high-pass filter is performed using a graphics processing unit (GPU), a tone-mapper processer, an image signal processor, or other image pre-processor. In some embodiments, the high-pass data component represents features and/or edges of the captured sensor data. In various embodiments, the high-pass filter is constructed to preserve the response of the first layer of a deep learning process. For example, a high-pass filter is constructed to preserve the response to a small filter at the top of a machine learning network. The relevant signal information for the first layer of the network is preserved such that the result of the analysis performed on a high-pass component data after the first layer is similar to the analysis performed on non-filtered data after the first layer. In various embodiments, the results are preserved for filters as small as a 5x5 matrix filter.

[0050] At 413, a low-pass filter is performed on the data. For example, a low -pass filter is performed on the captured sensor data to extract low -pass component data. In some embodiments, the low-pass filter is performed using a graphics processing unit (GPU), a tone-mapper processer, an image signal processor, or other image pre-processor. In some embodiments, the low-pass data component represents global data of the captured sensor data such as global illumination data.

[0051] In various embodiments, the filtering performed at 403 and 413 may use the same or different image pre-processors. For example, a tone-mapper processor is used to extract a high-pass data component and a graphics processing unit (GPU) is used to extract a low-pass data component. In some embodiments, the high-pass or low-pass data is extracted by subtracting one of the data components from the original captured data.

[0052] At 405 and 415, post-processing is performed on the respective high-pass and low-pass data components. In various embodiments, different post-processing techniques are utilized to enhance the signal quality and/or to reduce the amount of data required to represent the data. For example, a de-noising, demosaicing, local contrast enhancement, gain adjustment, and/or thresholding process, among others, may be performed on the respective high-pass and/or low-pass data components. In some embodiments, the data components are compressed and/or down-sampled. For example, once the high-pass and/or low-pass data is extracted, the respective data components may be compressed to more efficiently utilize the entire bit depth range. In some embodiments, the respective data components are compressed or quantized from a higher bit depth as captured by sensors to a lower bit depth compatible with the deep learning network. For example, a sensor data captured at 12-bits, 16-bits, 20-bits, 32-bits, or another appropriate bit depth per channel may be compressed or quantized to a lower bit depth such as 8-bits per channel. In

some embodiments, the post-processing steps at 405 and/or 415 are optional.

[0053] At 417, the low-pass data component is down-sampled. In various embodiments, the low-pass data component is fed into the network at a later stage of the network and may be down-sampled to a more efficient resource size. For example, a low -pass data component may be extracted at the full sensor size and reduced to a half or a quarter of the original size. Other percentages of reductions are possible as well. In various embodiments, the low-pass data is down-sampled but retains the relevant signal information. In many scenarios, the low-pass data can be easily down-sampled without losing signal information. By down-sampling the data, the data is more easily and quickly analyzed at a later layer in the deep learning network.

[0054] At 407, deep learning analysis is performed on the high-pass data component. In some embodiments, the high-pass data component is fed into the initial layer of the deep learning network and represents the most significant data for feature and edge detection. In various embodiments, the results of the deep learning analysis on the first layer using the high-pass data component are fed into subsequent layers of the network. For example, a neural network may include multiple layers, for example, five or more layers. The first layer receives the high-pass data component as input and the second layer receives the results of the deep learning analysis performed by the first layer. In various embodiments, the second or later layer receives the low-pass data components as additional input to perform additional deep learning analysis.

[0055] At 421, additional deep learning analysis is performed using the results of the analysis performed at 407 and the low-pass data component down-sampled at 417. In various embodiments, the deep learning analysis infers a vehicle control result. For example, the result of the deep learning analysis at 407 and 421 is used to control the vehicle for autonomous driving.

[0056] Figure 5 is a flow diagram illustrating an embodiment of a process for performing machine learning processing using high-pass, band-pass, and low-pass component data. In the example shown, the process of Figure 5 is used to extract three or more data components from sensor data and to provide the components at different layers of a deep learning network such as an artificial neural network. Similar to the process of Figure 4, a high-pass and low-pass component is extracted. In addition, the process of Figure 5 extracts one or more band-pass data components. In various embodiments, the decomposition of the sensor data into multiple components that are provided to different layers of the deep learning network allows the deep learning analysis to emphasize different sets of data at different layers of the network.

[0057] In some embodiments, the process of Figure 5 is used to perform the process of

Figures 1, 2, 3, and/or 4. In some embodiments, step 501 is performed at 103 of Figure 1, at 203 of Figure 2, at 303 of Figure 3, and/or at 401 of Figure 4. In some embodiments, step 503 is performed at 103 of Figure 1, at 205 of Figure 2, at 311 of Figure 3, and/or at 403 of Figure 4; step 513 is performed at 103 of Figure 1, at 205 of Figure 2, and/or at 311 or 321 of Figure 3; and/or step 523 is performed at 103 of Figure 1, at 205 of Figure 2, at 321 of Figure 3, and/or at 413 of Figure 4. In some embodiments, step 505 is performed at 103 of Figure 1, at 207 and 209 of Figure 2, at 313 of Figure 3, and/or at step 405 of Figure 4; step 515 is performed at 103 of Figure 1, at 207 and 209 of Figure 2, at 313, 323 and/or 325 of Figure 3, and/or at 405, 415, and/or 417 of Figure 4; and/or step 525 is performed at 103 of Figure 1, at 207 and 209 of Figure 2, at 323 and 325 of Figure 3, and/or at 415 and 417 of Figure 4. In some embodiments, step 537 is performed at 105 of Figure 1, at 211 of Figure 2, at 315 and 335 of Figure 3; and/or at 407 and 421 of Figure 4.

[0058] At 501, data is pre-processed. In some embodiments, the data is the sensor data captured from one or more sensors such as high dynamic range camera, radar, ultrasonic, and/or LiDAR sensors. In various embodiments, the data is pre-processed as described with respect to 103 of Figure 1, 203 of Figure 2, 303 of Figure 3, and/or 401 of Figure 4. Once the data is pre-processed, processing continues to 503, 513, and 523. In some embodiments, steps 503, 513, and 523 are run in parallel.

[0059] At 503, a high-pass filter is performed on the data. For example, a high-pass filter is performed on the captured sensor data to extract high-pass component data. In some embodiments, the high-pass filter is performed using a graphics processing unit (GPU), a tone-mapper processer, an image signal processor, or another image pre-processor. In some embodiments, the high-pass data component represents features and/or edges of the captured sensor data.

[0060] At 513, one or more band-pass filters are performed on the data to extract one or more band-pass data components. For example, a band-pass filter is performed on the captured sensor data to extract component data that includes a mix of feature, edge, intermediate, and/or global data. In various embodiments, one more band-pass components may be extracted. In some embodiments, the low-pass filter is performed using a graphics processing unit (GPU), a tone-mapper processer, an image signal processor, or another image pre-processor. In some embodiments, the band-pass data component represents data that is neither primarily edge/feature data nor primarily global data of the captured sensor data. In some embodiments, the band-pass data is utilized to preserve data fidelity that may be lost using only a high-pass data component and a low-pass data component.

[0061] At 523, a low-pass filter is performed on the data. For example, a low-pass filter is performed on the captured sensor data to extract low -pass component data. In some embodiments, the low-pass filter is performed using a graphics processing unit (GPU), a tone-mapper processer, an image signal processor, or other image pre-processor. In some embodiments, the low-pass data component represents global data of the captured sensor data such as global illumination data.

[0062] In various embodiments, the filtering performed at 503, 513, and 523 may use the same or different image pre-processors. For example, a tone-mapper processor is used to extract a high-pass data component and a graphics processing unit (GPU) is used to extract band-pass, and/or low-pass data components. In some embodiments, data components are extracted by subtracting one or more data components from the original captured data.

[0063] At 505, 515, and 525, post-processing is performed on the respective high-pass, band-pass, and low -pass data components. In various embodiments, different post- processing techniques are utilized to enhance the signal quality and/or to reduce the amount of data required to represent the data. In some embodiments, the different components are compressed and/or down-sampled to the appropriate size for the network layer receiving the data component. In various embodiments, the high-pass data will have a higher resolution than the band-pass data and the band-pass data will have a higher resolution than the low-pass data. In some embodiments, different band-pass data components will also have different resolutions as appropriate for the network layer each is provided as input for. In some embodiments, the respective data components are compressed or quantized from a higher bit depth as captured by sensors to a lower bit depth compatible with the deep learning network. For example, a sensor data captured at 12-bits per channel may be compressed or quantized to 8-bits per channel. In various embodiments, the pre processing filters are applied as described with respect to 207 of Figure 2 and/or 405, 415, and/or 417 of Figure 4.

[0064] At 537, deep learning analysis is performed using the data component results of 505,

515, and 525. In some embodiments, the high-pass data component is fed into the initial layer of the deep learning network and represents the most significant data for feature and edge detection. The one or more band-pass data components are fed into middle layer(s) of the network and include additional data for identifying features/edges and/or beneficial intermediate or global information. The low-pass data component is fed into a later layer of the network and includes global information to improve the analysis results of the deep learning network. In performing the deep learning analysis, additional data components representing different sensor data are fed into different layers as the analysis progresses to increase the accuracy of the result. In various

embodiments, the deep learning analysis infers a vehicle control result. For example, the result of the deep learning analysis is used to control the vehicle for autonomous driving. In some embodiments, a machine learning result is provided to a vehicle control module to at least in part autonomously operate a vehicle.

[0065] Figure 6 is a block diagram illustrating an embodiment of a deep learning system for autonomous driving. In some embodiments, the deep learning system of Figure 6 may be used to implement autonomous driving features for self-driving and driver-assisted automobiles. For example, using sensors affixed to a vehicle, sensor data is captured, processed as different input components, and fed into different stages of a deep learning network. The result of deep learning analysis is used by a vehicle control module to assist in the operation of the vehicle. In some embodiments, the vehicle control module is utilized for self-driving or driver-assisted operation of the vehicle. In various embodiments, the processes of Figures 1-5 utilize a deep learning system such as the one described in Figure 6.

[0066] In the example shown, deep learning system 600 is a deep learning network that includes sensors 601, image pre-processor 603, deep learning network 605, artificial intelligence (AI) processor 607, vehicle control module 609, and network interface 611. In various embodiments, the different components are communicatively connected. For example, sensor data from sensors 601 is fed to image pre-processor 603. Processed sensor data components of image pre-processor 603 are fed to deep learning network 605 running on AI processor 607. The output of deep learning network 605 running on AI processor 607 is fed to vehicle control module 609. In various embodiment, network interface 611 is used to communicate with remote servers, to make phone calls, to send and/or receive text messages, etc. based on the autonomous operation of the vehicle.

[0067] In some embodiments, sensors 601 include one or more sensors. In various embodiments, sensors 601 may be affixed to a vehicle, at different locations of the vehicle, and/or oriented in one or more different directions. For example, sensors 601 may be affixed to the front, sides, rear, and/or roof, etc. of the vehicle in forward-facing, rear-facing, side-facing, etc.

directions. In some embodiments, sensors 610 may be image sensors such as high dynamic range cameras. In some embodiments, sensors 601 include non-visual sensors. In some embodiments, sensors 601 include radar, LiDAR, and/or ultrasonic sensors, among others. In some embodiments, sensors 601 are not mounted to the vehicle with vehicle control module 609. For example, sensors 601 may be mounted on neighboring vehicles and/ or affixed to the road or environment and are included as part of a deep learning system for capturing sensor data.

[0068] In some embodiments, image pre-processor 603 is used to pre-process sensor data of sensors 601. For example, image pre-processor 603 may be used to pre-process the sensor data, split sensor data into one or more components, and/or post-process the one or more components. In some embodiments, image pre-processor 603 is a graphics processing unit (GPU), a central processing unit (CPU), an image signal processor, or a specialized image processor. In various embodiments, image pre-processor 603 is a tone-mapper processor to process high dynamic range data. In some embodiments, image pre-processor 603 is implemented as part of artificial intelligence (AI) processor 607. For example, image pre-processor 603 may be a component of AI processor 607.

[0069] In some embodiments, deep learning network 605 is a deep learning network for implementing autonomous vehicle control. For example, deep learning network 605 may be an artificial neural network such as a convolutional neural network (CNN) that is trained using sensor data and used to output vehicle control results to vehicle control module 609. In various embodiments, deep learning network 605 is a multi-stage learning network and can receive input data at two or more different stages of the network. For example, deep learning network 605 may receive feature and/or edge data at a first layer of deep learning network 605 and global data at a later layer (e.g., a second or third, etc. layer) of deep learning network 605. In various

embodiments, deep learning network 605 receives data at two or more different layers of the network and may compress and/or downsize the data as it is processed through different layers.

For example, the data size at layer one is a resolution that is higher than the data at a subsequent stage. In some embodiments, the data size at layer one is the full resolution of the captured image data and the data at a subsequent layer is a lower resolution (e.g., a quarter of the size) of the captured image data. In various embodiments, the input data received from image pre-processor 603 at subsequent layer(s) of deep learning network 605 matches the internal data resolution(s) of the data that is processed through the one or more previous layers.

[0070] In some embodiments, artificial intelligence (AI) processor 607 is a hardware processor for running deep learning network 605. In some embodiments, AI processor 607 is a specialized AI processor for performing inference using a convolutional neural network (CNN) on sensor data. In some embodiments, AI processor 607 is optimized for the bit depth of the sensor data. In some embodiments, AI processor 607 is optimized for deep learning operations such as neural network operations including convolution, dot-product, vector, and/or matrix operations, among others. In some embodiments, AI processor 607 is implemented using a graphics processing unit (GPU). In various embodiments, AI processor 607 is coupled to memory that is

configured to provide the AI processor with instructions which when executed cause the AI processor to perform deep learning analysis on the received input sensor data and to determine a machine learning result used to at least in part autonomously operate a vehicle.

[0071] In some embodiments, vehicle control module 609 is utilized to process the output of artificial intelligence (AI) processor 607 and to translate the output into a vehicle control operation. In some embodiments, vehicle control module 609 is utilized to control the vehicle for autonomous driving. In some embodiments, vehicle control module 609 can adjust the speed and/or steering of the vehicle. For example, vehicle control module 609 may be used to control a vehicle by braking, steering, changing lanes, accelerating and merging into another lane, etc. In some embodiments, vehicle control module 609 is used to control vehicle lighting such as brake lights, turns signals, headlights, etc. In some embodiments, vehicle control module 609 is used to control vehicle audio conditions such as the vehicle's sound system, playing audio alerts, enabling a microphone, enabling the horn, etc. In some embodiments, vehicle control module 609 is used to control notification systems including warning systems to inform the driver and/or passengers of driving events such as a potential collision or the approach of an intended destination. In some embodiments, vehicle control module 609 is used to adjust sensors such as sensors 601 of a vehicle. For example, vehicle control module 609 may be used to change parameters of one or more sensors such as modifying the orientation, changing the output resolution and/or format type, increasing or decreasing the capture rate, adjusting the captured dynamic range, adjusting the focus of a camera, enabling and/or disabling a sensor, etc. In some embodiments, vehicle control module 609 may be used to change parameters of image pre-processor 603 such as modifying the frequency range of filters, adjusting feature and/or edge detection parameters, adjusting channels and bit depth, etc. In various embodiments, vehicle control module 609 is used to implement self-driving and/or driver-assisted control of a vehicle.

[0072] In some embodiments, network interface 611 is a communication interface for sending and/or receiving data including voice data. In various embodiments, a network interface 611 includes a cellular or wireless interface for interfacing with remote servers, to connect and make voice calls, to send and/or receive text messages, etc. For example, network interface 611 may be used to receive an update for the instructions and/or operating parameters for sensors 601, image pre-processor 603, deep learning network 605, AI processor 607, and/or vehicle control module 609. For example, a machine learning model of deep learning network 605 may be updated using network interface 611. As another example, network interface 611 may be used to update firmware of sensors 601 and/or operating parameters of image pre-processor 603 such as image

processing parameters. In some embodiments, network interface 611 is used to make emergency contact with emergency services in the event of an accident or near-accident. For example, in the event of a collision, network interface 611 may be used to contact emergency services for help and may inform the emergency services of the location of the vehicle and collision details. In various embodiments, network interface 611 is used to implement autonomous driving features such as accessing calendar information to retrieve and/or update a destination location and/or expected arrival time.

[0073] Although the foregoing embodiments have been described in some detail for purposes of clarity of understanding, the invention is not limited to the details provided. There are many alternative ways of implementing the invention. The disclosed embodiments are illustrative and not restrictive.