Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V20/00—Scenes; Scene-specific elements
  • G06V20/50—Context or environment of the image
  • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
  • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F18/00—Pattern recognition
  • G06F18/20—Analysing
  • G06F18/25—Fusion techniques
  • G06F18/251—Fusion techniques of input or preprocessed data
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N20/00—Machine learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T7/00—Image analysis
  • G06T7/70—Determining position or orientation of objects or cameras
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V10/00—Arrangements for image or video recognition or understanding
  • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
  • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
  • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
  • G06V10/803—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of input or preprocessed data
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
  • G06V20/00—Scenes; Scene-specific elements
  • G06V20/50—Context or environment of the image
  • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
  • G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
  • G06V20/584—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads of vehicle lights or traffic lights
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/20—Special algorithmic details
  • G06T2207/20081—Training; Learning
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/30—Subject of image; Context of image processing
  • G06T2207/30248—Vehicle exterior or interior
  • G06T2207/30252—Vehicle exterior; Vicinity of vehicle
  • G06T2207/30261—Obstacle

Definitions

• Autonomous driving systems typically rely on mounting numerous sensors including a collection of vision and emitting distance sensors (e.g., radar sensor, lidar sensor, ultrasonic sensor, etc.) on a vehicle. The data captured by each sensor is then gathered to help understand the vehicle's surrounding environment and to determine how to control the vehicle. Vision sensors can be used to identify objects from captured image data and emitting distance sensors can be used to determine the distance of the detected objects. Steering and speed adjustments can be applied based on detected obstacles and clear drivable paths. But as the number and types of sensors increases, so does the complexity and cost of the system. For example, emitting distance sensors such as lidar are often costly to include in a mass market vehicle. Moreover, each additional sensor increases the input bandwidth requirements for the autonomous driving system. Therefore, there exists a need to find the optimal configuration of sensors on a vehicle. The configuration should limit the total number of sensors without limiting the amount and type of data captured to accurately describe the surrounding environment and safely control the vehicle.
• vision sensors e.g., radar sensor, lidar sensor
• One embodiment includes a system.
• the system comprises one or more processors configured to: receive image data based on an image captured using a camera of a vehicle; and utilize the image data as a basis of an input to a trained machine learning model to at least in part identify a distance of an object from the vehicle; wherein the trained machine learning model has been trained using a training image and a correlated output of an emitting distance sensor; and a memory coupled to the one or more processors.
• Another embodiment includes a computer program product, the computer program product being embodied in a non-transitory computer readable storage medium and comprising computer instructions.
• the computer instructions are for receiving image data based on an image captured using a camera of a vehicle; and utilizing the image data as a basis of an input to a trained machine learning model to at least in part identify a distance of an object from the vehicle, wherein the trained machine learning model has been trained using a training image and a correlated output of an emitting distance sensor.
• Yet another embodiment includes a method.
• the method comprises receiving a selected image based on an image captured using a camera of a vehicle; receiving distance data based on an emitting distance sensor of the vehicle; identifying an object using the selected image as an input to a trained machine learning model; extracting a distance estimate of the identified object from the received distance data; creating a training image by annotating the selected image with the extracted distance estimate; training a second machine learning model to predict a distance measurement using a training data set that includes the training image; and providing the trained second machine learning model to a second vehicle equipped with a second camera.
• Figure l is a block diagram illustrating an embodiment of a deep learning system for autonomous driving.
• Figure 2 is a flow diagram illustrating an embodiment of a process for creating training data for predicting object properties.
• Figure 3 is a flow diagram illustrating an embodiment of a process for training and applying a machine learning model for autonomous driving.
• Figure 4 is a flow diagram illustrating an embodiment of a process for training and applying a machine learning model for autonomous driving.
• Figure 5 is a diagram illustrating an example of capturing auxiliary sensor data for training a machine learning network.
• Figure 6 is a diagram illustrating an example of predicting object properties.
• 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.
• these implementations, or any other form that the invention may take, may be referred to as techniques.
• the order of the steps of disclosed processes may be altered within the scope of the invention.
• 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.
• the term‘processor’ refers to one or more devices, circuits, and/or processing cores configured to process data, such as computer program instructions.
• a machine learning training technique for generating highly accurate machine learning results from vision data is disclosed.
• auxiliary sensor data such as radar and lidar results
• the auxiliary data is associated with objects identified from the vision data to accurately estimate object properties such as object distance.
• the collection and association of auxiliary data with vision data is done automatically and requires little, if any, human intervention.
• objects identified using vision techniques do not need to be manually labeled, significantly improving the efficiency of machine learning training.
• the training data can be automatically generated and used to train a machine learning model to predict object properties with a high degree of accuracy.
• the data may be collected automatically from a fleet of vehicles by collecting snapshots of the vision data and associated related data, such as radar data.
• only a subset of the vision-radar related association targets are sampled.
• the collected fusion data from the fleet of vehicles is automatically collected and used to train neural nets to mimic the captured data.
• the trained machine learning model can be deployed to vehicles for accurately predicting object properties, such as distance, direction, and velocity, using only vision data. For example, once the machine learning model has been trained to be able to determine an object distance using images of a camera without a need of a dedicated distance sensor, it may become no longer necessary to include a dedicated distance sensor in an autonomous driving vehicle. When used in conjunction with a dedicated distance sensor, this machine learning model can be used as a redundant or a secondary distance data source to improve accuracy and/or provide fault tolerance.
• the identified objects and corresponding properties can be used to implement autonomous driving features such as self-driving or driver-assisted operation of a vehicle. For example, an autonomous vehicle can be controlled to avoid a merging vehicle identified using the disclosed techniques.
• a system comprising one or more processors coupled to memory is configured to receive image data based on an image captured using a camera of a vehicle.
• a processor such as an artificial intelligence (AI) processor installed on an autonomous vehicle receives image data from a camera, such as a forward-facing camera of the vehicle. Additional cameras such as side-facing and rear-facing cameras can be used as well.
• the image data is utilized as a basis of an input to a machine learning trained model to at least in part identify a distance of an object from the vehicle.
• the captured image is used as an input to a machine learning model such as a model of a deep learning network running on the AI processor.
• the model is used to predict the distance of objects identified in the image data.
• the trained machine learning model has been trained using a training image and a correlated output of an emitting distance sensor.
• Emitting distance sensors may emit a signal (e.g., radio signal, ultrasonic signal, light signal, etc.) in detecting a distance of an object from the sensor.
• a radar sensor mounted to a vehicle emits radar to identify the distance and direction of surrounding obstacles.
• the distances are then correlated to objects identified in a training image captured from the vehicle's camera.
• the associated training image is annotated with the distance measurements and used to train a machine learning model.
• the model is used to predict additional properties such as an object's velocity.
• the velocity of objects determined by radar is associated with objects in the training image to train a machine learning model to predict object velocities and directions.
• a vehicle is equipped with sensors to capture the environment of the vehicle and vehicle operating parameters.
• the captured data includes vision data (such as video and/or still images) and additional auxiliary data such as radar, lidar, inertia, audio, odometry, location, and/or other forms of sensor data.
• the sensor data may capture vehicles, pedestrians, vehicle lane lines, vehicle traffic, obstacles, traffic control signs, traffic sounds, etc.
• Odometry and other similar sensors capture vehicle operating parameters such as vehicle speed, steering, orientation, change in direction, change in location, change in elevation, change in speed, etc.
• the captured vision and auxiliary data is transmitted from the vehicle to a training server for creating a training data set.
• the transmitted vision and auxiliary data is correlated and used to automatically generate training data.
• the training data is used to train a machine learning model for generating highly accurate machine learning results.
• a time series of captured data is used to generate the training data.
• a ground truth is determined based on a group of time series elements and is used to annotate at least one of the elements, such as a single image, from the group. For example, a series of images and radar data for a time period, such as 30 seconds, are captured.
• a vehicle identified from the image data and tracked across the time series is associated with a corresponding radar distance and direction from the time series.
• the associated auxiliary data such as radar distance data, is associated with the vehicle after analyzing the image and distance data captured for the time series.
• ambiguities such as multiple objects with similar distances can be resolved with a high degree of accuracy to determine a ground truth. For example, when using only a single captured image, there may be insufficient corresponding radar data to accurately estimate the different distances of two cars in the event one car occludes another or when two cars are close together. By tracking the cars over a time series, however, the distances identified by radar can be properly associated with the correct cars as the cars separate, travel in different directions, and/or travel at different speeds, etc.
• one or more images of the time series are converted to training images and annotated with the corresponding ground truth such as the distance, velocity, and/or other appropriate object properties.
• a machine learning model trained using auxiliary sensor data can accurately predict the result of an auxiliary sensor without the need for the physical auxiliary sensor.
• training vehicles can be equipped with auxiliary sensors, including expensive and/or difficult to operate sensors, for collecting training data.
• the training data can then be used to train a machine learning model for predicting the result of an auxiliary sensor, such as a radar, lidar, or another sensor.
• the trained model is then deployed to vehicles, such as production vehicles, that only require vision sensors.
• the auxiliary sensors are not required but can be used as a secondary data source.
• the trained model is used in the case of auxiliary sensors failing.
• the trained machine learning model uses input from one or more vision sensors to predict the result of the auxiliary sensors.
• the predicted results can be used for implementing autonomous driving features that require detecting objects (e.g., pedestrians, stationary vehicles, moving vehicles, curbs, obstacles, road barriers, etc.) and their distance and direction.
• the predicted results can be used to detect the distance and direction of traffic control objects such as traffic lights, traffic signs, street signs, etc.
• vision sensors and object distance are used in the previous examples, alternative sensors and predicted properties are possible as well.
• FIG. 1 is a block diagram illustrating an embodiment of a deep learning system for autonomous driving.
• the deep learning system includes different components that may be used together for self-driving and/or driver-assisted operation of a vehicle as well as for gathering and processing data for training a machine learning model.
• the deep learning system is installed on a vehicle and data captured from the vehicle can be used to train and improve the deep learning system of the vehicle or other similar vehicles.
• the deep learning system may be used to implement autonomous driving functionality including identifying objects and predicting object properties such as distance and direction using vision data as input.
• deep learning system 100 is a deep learning network that includes vision sensors 101, additional sensors 103, image pre-processor 105, deep learning network 107, artificial intelligence (AI) processor 109, vehicle control module 111, and network interface 113.
• the different components are communicatively connected.
• image data captured from vision sensors 101 is fed to image pre processor 105.
• Processed sensor data of image pre-processor 105 is fed to deep learning network 107 running on AI processor 109.
• sensor data from additional sensors 103 is used as an input to deep learning network 107.
• the output of deep learning network 107 running on AI processor 109 is fed to vehicle control module 111.
• vehicle control module 111 is connected to and controls the operation of the vehicle such as the speed, braking, and/or steering, etc. of the vehicle.
• sensor data and/or machine learning results can be sent to a remote server (not shown) via network interface 113.
• sensor data such as data captured from vision sensors 101 and/or additional sensors 103, can be transmitted to a remote training server via network interface 113 to collect training data for improving the performance, comfort, and/or safety of the vehicle.
• network interface 113 is used to communicate with remote servers, to make phone calls, to send and/or receive text messages, and to transmit sensor data based on the operation of the vehicle, among other reasons.
• deep learning system 100 may include additional or fewer components as appropriate.
• image pre processor 105 is an optional component.
• a post processing component (not shown) is used to perform post-processing on the output of deep learning network 107 before the output is provided to vehicle control module 111.
• vision sensors 101 include one or more camera sensors for capturing image data.
• vision sensors 101 may be affixed to a vehicle, at different locations of the vehicle, and/or oriented in one or more different directions.
• vision sensors 101 may be affixed to the front, sides, rear, and/or roof, etc. of the vehicle in forward-facing, rear-facing, side-facing, etc. directions.
• vision sensors 101 may be image sensors such as high dynamic range cameras and/or cameras with different fields of view.
• 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.
• 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.
• vision sensors 101 are not mounted to the vehicle with vehicle control module 111.
• vision sensors 101 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.
• vision sensors 101 include one or more cameras that capture the surrounding environment of the vehicle, including the road the vehicle is traveling on.
• one or more front-facing and/or pillar cameras capture images of objects such as vehicles, pedestrians, traffic control objects, roads, curbs, obstacles, etc. in the environment surrounding the vehicle.
• cameras capture a time series of image data including image data of neighboring vehicles including those attempting to cut into the lane the vehicle is traveling in.
• Vision sensors 101 may include image sensors capable of capturing still images and/or video.
• the data may be captured over a period of time, such as a sequence of captured data over a period of time, and synchronized with other vehicle data including other sensor data.
• image data used to identify objects may be captured along with radar and odometry data over a period of 15 seconds or another appropriate period.
• additional sensors 103 include additional sensors for capturing sensor data in addition to vision sensors 101.
• additional sensors 103 may be affixed to a vehicle, at different locations of the vehicle, and/or oriented in one or more different directions.
• additional sensors 103 may be affixed to the front, sides, rear, and/or roof, etc. of the vehicle in forward-facing, rear-facing, side-facing, etc. directions.
• additional sensors 103 may be emitting sensors such as radar, ultrasonic, and/or lidar sensors.
• additional sensors 103 include non-visual sensors.
• Additional sensors 103 may include radar, audio, lidar, inertia, odometry, location, and/or ultrasonic sensors, among others.
• twelve ultrasonic sensors may be affixed to the vehicle to detect both hard and soft objects.
• a forward-facing radar is utilized to capture data of the surrounding environment.
• radar sensors are able to capture surrounding detail despite heavy rain, fog, dust, and other vehicles.
• additional sensors 103 are not mounted to the vehicle with vehicle control module 111.
• additional sensors 103 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.
• additional sensors 103 include one or more sensors that capture the surrounding environment of the vehicle, including the road the vehicle is traveling on.
• a forward-facing radar sensor captures the distance data of objects in the forward field of view of the vehicle.
• Additional sensors may capture odometry, location, and/or vehicle control information including information related to vehicle trajectory.
• Sensor data may be captured over a period of time, such as a sequence of captured data over a period of time, and associated with image data captured from vision sensors 101.
• additional sensors 103 include location sensors such as global position system (GPS) sensors for determining the vehicle's location and/or change in location.
• GPS global position system
• one or more sensors of additional sensors 103 are optional and are included only on vehicles designed for capturing training data.
• Vehicles without one or more sensors of additional sensors 103 can simulate the results of additional sensors 103 by predicting the output using a trained machine learning model and the techniques disclosed herein. For example, vehicles without a forward-facing radar or lidar sensor can predict the results of the optional sensor using image data by applying a trained machine learning model, such as the model of deep learning network 107.
• image pre-processor 105 is used to pre-process sensor data of vision sensors 101.
• image pre-processor 105 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.
• image pre-processor 105 is a graphics processing unit (GPU), a central processing unit (CPU), an image signal processor, or a specialized image processor.
• image pre-processor 105 is a tone-mapper processor to process high dynamic range data.
• image pre-processor 105 is implemented as part of artificial intelligence (AI) processor 109.
• AI artificial intelligence
• image pre processor 105 may be a component of AI processor 109.
• image pre processor 105 may be used to normalize an image or to transform an image.
• image pre-processor 105 may be used to transform the image to remove or modify the warping.
• noise, distortion, and/or blurriness is removed or reduced during a pre-processing step.
• the image is adjusted or normalized to improve the result of machine learning analysis. For example, the white balance of the image is adjusted to account for different lighting operating conditions such as daylight, sunny, cloudy, dusk, sunrise, sunset, and night conditions, among others.
• deep learning network 107 is a deep learning network used for determining vehicle control parameters including analyzing the driving environment to determine objects and their corresponding properties such as distance, velocity, or another appropriate parameter.
• deep learning network 107 may be an artificial neural network such as a convolutional neural network (CNN) that is trained on input such as sensor data and its output is provided to vehicle control module 111.
• CNN convolutional neural network
• the output may include at least a distance estimate of detected objects.
• the output may include at least potential vehicles that are likely to merge into the vehicle's lane, their distances, and their velocities.
• deep learning network 107 receives as input at least image sensor data, identifies objects in the image sensor data, and predicts the distance of the objects. Additional input may include scene data describing the environment around the vehicle and/or vehicle specifications such as operating characteristics of the vehicle. Scene data may include scene tags describing the environment around the vehicle, such as raining, wet roads, snowing, muddy, high density traffic, highway, urban, school zone, etc.
• the output of deep learning network 107 is a three-dimensional representation of a vehicle's surrounding environment including cuboids representing objects such as identified objects.
• the output of deep learning network 107 is used for autonomous driving including navigating a vehicle towards a target destination.
• AI processor 109 is a hardware processor for running deep learning network 107.
• AI processor 109 is a specialized AI processor for performing inference using a convolutional neural network (CNN) on sensor data.
• CNN convolutional neural network
• AI processor 109 may be optimized for the bit depth of the sensor data.
• AI processor 109 is optimized for deep learning operations such as neural network operations including convolution, dot-product, vector, and/or matrix operations, among others.
• AI processor 109 is implemented using a graphics processing unit (GPU).
• GPU graphics processing unit
• AI processor 109 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, such as an object distance, used for autonomous driving.
• AI processor 109 is used to process sensor data in preparation for making the data available as training data.
• vehicle control module 111 is utilized to process the output of artificial intelligence (AI) processor 109 and to translate the output into a vehicle control operation.
• vehicle control module 111 is utilized to control the vehicle for autonomous driving.
• vehicle control module 111 can adjust speed, acceleration, steering, braking, etc. of the vehicle.
• vehicle control module 111 is used to control the vehicle to maintain the vehicle's position within a lane, to merge the vehicle into another lane, to adjust the vehicle's speed and lane positioning to account for merging vehicles, etc.
• vehicle control module 111 is used to control vehicle lighting such as brake lights, turns signals, headlights, etc.
• vehicle control module 111 is used to control vehicle audio conditions such as the vehicle's sound system, playing audio alerts, enabling a microphone, enabling the horn, etc.
• vehicle control module 111 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.
• vehicle control module 111 is used to adjust sensors such as vision sensors 101 and additional sensors 103 of a vehicle.
• vehicle control module 111 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.
• vehicle control module 111 may be used to change parameters of image pre-processor 105 such as modifying the frequency range of filters, adjusting feature and/or edge detection parameters, adjusting channels and bit depth, etc.
• vehicle control module 111 is used to implement self-driving and/or driver-assisted control of a vehicle.
• vehicle control module 111 is implemented using a processor coupled with memory.
• vehicle control module 111 is implemented using an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or other appropriate processing hardware.
• ASIC application-specific integrated circuit
• PLD programmable logic device
• network interface 113 is a communication interface for sending and/or receiving data including training data.
• a network interface 113 includes a cellular or wireless interface for interfacing with remote servers, to transmit sensor data, to transmit potential training data, to receive updates to the deep learning network including updated machine learning models, to connect and make voice calls, to send and/or receive text messages, etc.
• network interface 113 may be used to transmit sensor data captured for use as potential training data to a remote training server for training a machine learning model.
• network interface 113 may be used to receive an update for the instructions and/or operating parameters for vision sensors 101, additional sensors 103, image pre-processor 105, deep learning network 107, AI processor 109, and/or vehicle control module 111.
• a machine learning model of deep learning network 107 may be updated using network interface 113.
• network interface 113 may be used to update firmware of vision sensors 101 and additional sensors 103 and/or operating parameters of image pre-processor 105 such as image processing parameters.
• Figure 2 is a flow diagram illustrating an embodiment of a process for creating training data for predicting object properties.
• image data is annotated with sensor data from additional auxiliary sensors to automatically create training data.
• a time series of elements made up of sensor and related auxiliary data is collected from a vehicle and used to automatically create training data.
• the process of Figure 2 is used to automatically label training data with corresponding ground truths.
• the ground truth and image data are packaged as training data to predict properties of objects identified from the image data.
• the sensor and related auxiliary data are captured using the deep learning system of Figure 1.
• the sensor data is captured from vision sensors 101 of Figure 1 and related data is captured from additional sensors 103 of Figure 1.
• the process of Figure 2 is performed to automatically collect data when existing predictions are incorrect or can be improved.
• a prediction is made by an autonomous vehicle to determine one or more object properties, such as distance and direction, from vision data.
• the prediction is compared to distance data received from an emitting distance sensor.
• a determination can be made whether the prediction is within an acceptable accuracy threshold.
• a determination is made that the prediction can be improved.
• the process of Figure 2 can be applied to the prediction scenario to create a curated set of training examples for improving the machine learning model.
• vision data is received.
• the vision data may be image data such as video and/or still images.
• the vision data is captured at a vehicle and transmitted to a training server.
• the vision data may be captured over a period of time to create a time series of elements.
• the elements include timestamps to maintain an ordering of the elements.
• objects in the time series can be tracked across the time series to better disambiguate objects that are difficult to identify from a single input sample, such as a single input image and corresponding related data. For example, a pair of oncoming headlights may appear at first to both belong to a single vehicle but in the event the headlights separate, each headlight is identified as belonging to a separate motorcycle.
• objects in the image data are easier to distinguish than objects in the auxiliary related data received at 203.
• the correct distance data can be associated with the identified van.
• sensor data captured as a time series is captured in the format that a machine learning model uses as input.
• the sensor data may be raw or processed image data.
• the time series in the event a time series of data is received, the time series may be organized by associating a timestamp with each element of the time series. For example, a timestamp is associated with at least the first element in a time series. The timestamp may be used to calibrate time series elements with related data such as data received at 203.
• the length of the time series may be a fixed length of time, such as 10 seconds, 30 seconds, or another appropriate length. The length of time may be configurable.
• the time series may be based on the speed of the vehicle, such as the average speed of the vehicle.
• the length of time for a time series may be increased to capture data over a longer distance traveled than would be possible if using a shorter time length for the same speed.
• the number of elements in the time series is configurable. The number of elements may be based on the distance traveled. For example, for a fixed time period, a faster moving vehicle includes more elements in the time series than a slower moving vehicle. The additional elements increase the fidelity of the captured environment and can improve the accuracy of the predicted machine learning results. In various embodiments, the number of elements is adjusted by adjusting the frames per second a sensor captures data and/or by discarding unneeded intermediate frames.
• the related data is received at a training server along with the vision data received at 201.
• the related data is sensor data from additional sensors of the vehicle, such as ultrasonic, radar, lidar, or other appropriate sensors.
• the related data may be distance, direction, velocity, location, orientation, change in location, change in orientation, and/or other related data captured by the vehicle's additional sensors.
• the related data may be used to determine a ground truth for features identified in the vision data received at 201. For example, distance and direction measurements from radar sensors are used to determine object distances and directions for objects identified in the vision data.
• the related data received is a time series of data corresponding to a time series of vision data received at 201.
• the data related to the vision data includes map data.
• offline data such as road and/or satellite level map data may be received at 203.
• the map data may be used to identify features such as roads, vehicle lanes, intersections, speed limits, school zones, etc.
• the map data can describe the path of vehicle lanes.
• estimated location of identified vehicles in vehicles lanes estimated distances for the detected vehicles can be determined/corroborated.
• the map data can describe the speed limit associated with different roads of the map.
• the speed limit data may be used to validate velocity vectors of identified vehicles.
• objects in the vision data are identified.
• the vision data is used as an input to identify objects in the surrounding environment of the vehicle. For example, vehicles, pedestrians, obstacles, etc. are identified from the vision data.
• the objects are identified using a deep learning system with a trained machine learning model.
• bounding boxes are created for identified objects. The bounding boxes may be two-dimensional bounding boxes or three-dimensional bounding boxes, such as cuboids, that outline the exterior of the identified object.
• additional data is used to help identify the objects, such as the data received at 203. The additional data may be used to increase the accuracy in object identification.
• a ground truth is determined for identified objects.
• ground truths are determined for the object identified at 205 from the vision data received at 201.
• the related data is depth (and/or distance) data of the identified objects.
• a machine learning model can be trained to estimate object distances by using the related distance data as the ground truth for detected objects.
• the distances are for detected objects such as an obstacle, a barrier, a moving vehicle, a stationary vehicle, traffic control signals, pedestrians, etc. and used as the ground truth for training.
• the ground truth for other object parameters such as direction, velocity, acceleration, etc. may be determined.
• accurate distances and directions are determined as ground truths for identified objects.
• accurate velocity vectors are determined as ground truths for identified objects, such as vehicles and pedestrians.
• vision data and related data are organized by timestamps and corresponding timestamps are used to synchronize the two data sets.
• timestamps are used to synchronize a time series of data, such as a sequence of images and a corresponding sequence of related data.
• the data may be synchronized at capture time. For example, as each element of a time series is captured, a corresponding set of related data is captured and saved with the time series element.
• the time period of the related data is configurable and/or matches the time period of the time series of elements.
• the related data is sampled at the same rate as the time series elements.
• the ground truth be determined. For example, analysis of only a subset of vision data may misidentify objects and/or their properties. By expanding the analysis across the entire time series, ambiguities are removed. For example, an occluded vehicle may be revealed earlier or later in the time series. Once identified, the sometimes-occluded vehicle can be tracked throughout the entire time series, even when occluded. Similarly, object properties for the sometimes-occluded vehicle can be tracked throughout the time series by associating the object properties from the related data to the identified object in the vision data. In some embodiments, the data is played backwards (and/or forwards) to determine any points of ambiguity when associating related data to vision data. The objects at different times in the time series may be used to help determine object properties for the objects across the entire time series.
• a threshold value is used to determine whether to associate an object property as a ground truth of an identified object.
• related data with a high degree of certainty is associated with an identified object while related data with a degree of certainty below a threshold value is not associated with the identified object.
• the related data may be conflicting sensor data.
• ultrasonic and radar data output may conflict.
• distance data may conflict with map data. The distance data may estimate a school zone begins in 30 meters while information from map data may describe the same school zone as starting in 20 meters. In the event the related data has a low degree of certainty, the related data may be discarded and not used to determine the ground truth.
• the ground truth is determined to predict semantic labels.
• a detected vehicle can be labeled based on a predicted distance and direction as being in the left lane or right lane.
• the detected vehicle can be labeled as being in a blind spot, as a vehicle that should be yielded to, or with another appropriate semantic label.
• vehicles are assigned to roads or lanes in a map based on the determined ground truth.
• the determined ground truth can be used to label traffic lights, lanes, drivable space, or other features that assist autonomous driving.
• the training data is packaged. For example, an element of vision data received at 201 is selected and associated with the ground truth determined at 207.
• the element selected is an element of a time series.
• the selected element represents sensor data input, such as a training image, to a machine learning model and the ground truth represents the predicted result.
• the selected data is annotated and prepared as training data.
• the training data is packaged into training, validation, and testing data. Based on the determined ground truth and selected training element, the training data is packaged to train a machine learning model to predict the results related to one or more related auxiliary sensors.
• the trained model can be used to accurately predict distances and directions of objects with results similar to measurements using sensors such as radar or lidar sensors.
• the machine learning results are used to implement features for autonomous driving.
• the packaged training data is now available for training a machine learning model.
• Figure 3 is a flow diagram illustrating an embodiment of a process for training and applying a machine learning model for autonomous driving.
• input data including a primary and secondary sensor data is received and processed to create training data for training a machine learning model.
• the primary sensor data corresponds to image data captured via an autonomous driving system and the secondary sensor data corresponds to sensor data captured from an emitting distance sensor.
• the secondary sensor data may be used to annotate the primary sensor data to train a machine learning model to predict an output based on the secondary sensor.
• the sensor data corresponds to sensor data captured based on particular use cases, such as the user manually disengaging autonomous driving or where distance estimates from vision data vary significantly from distance estimates from secondary sensors.
• the primary sensor data is sensor data of vision sensors 101 of Figure 1 and the secondary sensor data is sensor data of one or more sensors of additional sensors 103 of Figure 1.
• the process is used to create and deploy a machine learning model for deep learning system 100 of Figure 1.
• training data is prepared.
• sensor data including image data and auxiliary data is received to create a training data set.
• the image data may include still images and/or video from one or more cameras. Additional sensors such as radar, lidar, ultrasonic, etc. may be used to provide relevant auxiliary sensor data.
• the image data is paired with corresponding auxiliary data to help identify the properties of objects detected in the sensor data. For example, distance and/or velocity data from auxiliary data can be used to accurately estimate the distance and/or velocity of objects identified in the image data.
• the sensor data is a time series of elements and is used to determine a ground truth.
• the ground truth of the group is then associated with a subset of the time series, such as a frame of image data.
• the selected element of the time series and the ground truth are used to prepare the training data.
• the training data is prepared to train a machine learning model to only estimate properties of objects identified in the image data, such as the distance and direction of vehicles, pedestrians, obstacles, etc.
• the prepared training data may include data for training, validation, and testing.
• the sensor data may be of different formats. For example, sensor data may be still image data, video data, radar data, ultrasonic data, audio data, location data, odometry data, etc.
• the odometry data may include vehicle operation parameters such as applied acceleration, applied braking, applied steering, vehicle location, vehicle orientation, the change in vehicle location, the change in vehicle orientation, etc.
• the training data is curated and annotated for creating a training data set.
• a portion of the preparation of the training data may be performed by a human curator.
• a portion of the training data is generated automatically from data captured from vehicles, greatly reducing the effort and time required to build a robust training data set.
• the format of the data is compatible with a machine learning model used on a deployed deep learning application.
• the training data includes validation data for testing the accuracy of the trained model.
• the process of Figure 2 is performed at 301 of Figure 3.
• a machine learning model is trained.
• a machine learning model is trained using the data prepared at 301.
• the model is a neural network such as a convolutional neural network (CNN).
• CNN convolutional neural network
• the model includes multiple intermediate layers.
• the neural network may include multiple layers including multiple convolution and pooling layers.
• the training model is validated using a validation data set created from the received sensor data.
• the machine learning model is trained to predict an output of a sensor such as a distance emitting sensor from a single input image. For example, a distance and direction property of an object can be inferred from an image captured from a camera. As another example, a velocity vector of a neighboring vehicle including whether the vehicle will attempt to merge is predicted from an image captured from a camera.
• the trained machine learning model is deployed.
• the trained machine learning model is installed on a vehicle as an update for a deep learning network, such as deep learning network 107 of Figure 1.
• an over-the-air update is used to install the newly trained machine learning model.
• an over-the-air update can be received via a network interface of the vehicle such as network interface 113 of Figure 1.
• the update is a firmware update transmitted using a wireless network such as a WiFi or cellular network.
• the new machine learning model may be installed when the vehicle is serviced.
• sensor data is received.
• sensor data is captured from one or more sensors of the vehicle.
• the sensors are vision sensors 101 of Figure 1.
• the sensors may include image sensors such as a fisheye camera mounted behind a windshield, forward or side-facing cameras mounted in the pillars, rear-facing cameras, etc.
• the sensor data is in the format or is converted into a format that the machine learning model trained at 303 utilizes as input.
• the sensor data may be raw or processed image data.
• the sensor data is preprocessed using an image pre-processor such as image pre-processor 105 of Figure 1 during a pre-processing step.
• the image may be normalized to remove distortion, noise, etc.
• the received sensor data is data captured from ultrasonic sensors, radar, LiDAR sensors, microphones, or other appropriate technology and used as the expected input to the trained machine learning model deployed at 305.
• the trained machine learning model is applied.
• the machine learning model trained at 303 is applied to sensor data received at 307.
• the application of the model is performed by an AI processor such as AI processor 109 of Figure 1 using a deep learning network such as deep learning network 107 of Figure 1.
• one or more object properties such as an object distance, direction, and/or velocity are predicted from image data. For example, different objects are identified in the image data and an object distance and direction for each identified object are inferred using the trained machine learning model. As another example, a velocity vector of a vehicle is inferred for a vehicle identified in the image data.
• the velocity vector may be used to determine whether the neighboring vehicle is likely to cut into the current lane and/or the likelihood the vehicle is a safety risk.
• vehicles, pedestrians, obstacles, lanes, traffic control signals, map features, speed limits, drivable space, etc. and their related properties are identified by applying the machine learning model.
• the features are identified in three-dimensions, such as a three-dimensional velocity vector.
• the autonomous vehicle is controlled.
• one or more autonomous driving features are implemented by controlling various aspects of the vehicle. Examples may include controlling the steering, speed, acceleration, and/or braking of the vehicle, maintaining the vehicle's position in a lane, maintaining the vehicle's position relative to other vehicles and/or obstacles, providing a notification or warning to the occupants, etc.
• a vehicle's steering and speed may be controlled to maintain the vehicle safely between two lane lines and at a safe distance from other objects. For example, distances and directions of neighboring objects are predicted and a corresponding drivable space and driving path is identified.
• a vehicle control module such as vehicle control module 111 of Figure 1 controls the vehicle.
• Figure 4 is a flow diagram illustrating an embodiment of a process for training and applying a machine learning model for autonomous driving.
• the process of Figure 4 is utilized to collect and retain sensor data for training a machine learning model for autonomous driving.
• the process of Figure 4 is implemented on a vehicle enabled with autonomous driving whether the autonomous driving control is enabled or not.
• sensor data can be collected in the moments immediately after autonomous driving is disengaged, while a vehicle is being driven by a human driver, and/or while the vehicle is being autonomously driven.
• the techniques described by Figure 4 are implemented using the deep learning system of Figure 1.
• portions of the process of Figure 4 are performed at 307, 309, and/or 311 of Figure 3 as part of the process of applying a machine learning model for autonomous driving.
• sensor data is received.
• a vehicle equipped with sensors captures sensor data and provides the sensor data to a neural network running on the vehicle.
• the sensor data may be vision data, ultrasonic data, radar data, LiDAR data, or other appropriate sensor data.
• an image is captured from a high dynamic range forward-facing camera.
• ultrasonic data is captured from a side-facing ultrasonic sensor.
• a vehicle is affixed with multiple sensors for capturing data.
• 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.
• 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.
• ultrasonic and/or 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.
• the captured data from different sensors is associated with captured metadata to allow the data captured from different sensors to be associated together.
• the direction, field of view, frame rate, resolution, timestamp, and/or other captured metadata is received with the sensor data.
• the sensor data includes odometry data including the location, orientation, change in location, and/or change in orientation, etc. of the vehicle.
• location data is captured and associated with other sensor data captured during the same time frame.
• the location data captured at the time that image data is captured is used to associate location information with the image data.
• the received sensor data is provided for deep learning analysis.
• the sensor data is pre-processed.
• one or more pre processing passes may be performed on the sensor data.
• the data may be pre- processed to remove noise, to correct for alignment issues and/or blurring, etc.
• one or more different filtering passes are performed on the data.
• a high-pass filter may be performed on the data and a low-pass filter may be performed on the data to separate out different components of the sensor data.
• the pre processing step performed at 403 is optional and/or may be incorporated into the neural network.
• the deep learning analysis is initiated.
• the deep learning analysis is performed on the sensor data received at 401 and optionally pre-processed at 403.
• the deep learning analysis is performed using a neural network such as a convolutional neural network (CNN).
• CNN convolutional neural network
• the machine learning model is trained offline using the process of Figure 3 and deployed onto the vehicle for performing inference on the sensor data.
• the model may be trained to predict object properties such as distance, direction, and/or velocity.
• the model is trained to identify pedestrians, moving vehicles, parked vehicles, obstacles, road lane lines, drivable space, etc., as appropriate.
• a bounding box is determined for each identified object in the image data and a distance and direction is predicted for each identified object.
• the bounding box is a three- dimensional bounding box such as a cuboid.
• the bounding box outlines the exterior surface of the identified object and may be adjusted based on the size of the object. For example, different sized vehicles are represented using different sized bounding boxes (or cuboids).
• the object properties estimated by the deep learning analysis are compared to properties measured by sensors and received as sensor data.
• the neural network includes multiple layers including one or more intermediate layers and/or one or more different neural networks are utilized to analyze the sensor data.
• the sensor data and/or the results of deep learning analysis are retained and transmitted at 411 for the automatic generation of training data.
• the deep learning analysis is used to predict additional features.
• the predicted features may be used to assist autonomous driving.
• a detected vehicle can be assigned to a lane or road.
• a detected vehicle can be determined to be in a blind spot, to be a vehicle that should be yielded to, to be a vehicle in the left adjacent lane, to be a vehicle in the right adjacent lane, or to have another appropriate attribute.
• the deep learning analysis can identify traffic lights, drivable space, pedestrians, obstacles, or other appropriate features for driving.
• the results of deep learning analysis are provided to vehicle control.
• the results are provided to a vehicle control module to control the vehicle for autonomous driving and/or to implement autonomous driving functionality.
• the results of deep learning analysis at 405 are passed through one or more additional deep learning passes using one or more different machine learning models. For example, identified objects and their properties (e.g., distance, direction, etc.) may be used to determine drivable space. The drivable space is then used to determine a drivable path for the vehicle. Similarly, in some embodiments, a predicted vehicle velocity vector is detected. The determined path for the vehicle based at least in part on a predicted velocity vector is used to predict cut-ins and to avoid potential collisions.
• the various outputs of deep learning are used to construct a three-dimensional representation of the vehicle's environment for autonomous driving which includes identified objects, the distance and direction of identified objects, predicted paths of vehicles, identified traffic control signals including speed limits, obstacles to avoid, road conditions, etc.
• the vehicle control module utilizes the determined results to control the vehicle along a determined path.
• the vehicle control module is vehicle control module 111 of Figure 1. [0057] At 409, the vehicle is controlled. In some embodiments, a vehicle with autonomous driving activated is controlled using a vehicle control module such as vehicle control module 111 of Figure 1.
• the vehicle control can modulate the speed and/or steering of the vehicle, for example, to maintain a vehicle at a safe distance from other vehicles and in a lane at an appropriate speed in consideration of the environment around it.
• the results are used to adjust the vehicle in anticipation that a neighboring vehicle will merge into the same lane.
• a vehicle control module uses the results of deep learning analysis, determines the appropriate manner to operate the vehicle, for example, along a determined path with the appropriate speed.
• the result of vehicle controls such as a change in speed, application of braking, adjustment to steering, etc. are retained and used for the automatic generation of training data.
• the vehicle control parameters may be retained and transmitted at 411 for the automatic generation of training data.
• sensor and related data are transmitted.
• the sensor data received at 401 along with the results of deep learning analysis at 405 and/or vehicle control parameters used at 409 are transmitted to a computer server for the automatic generation of training data.
• the data is a time series of data and the various gathered data are associated together by a remote training computer server.
• image data is associated with auxiliary sensor data, such as distance, direction, and/or velocity data, to generate a ground truth.
• the collected data is transmitted wirelessly, for example, via a WiFi or cellular connection, from a vehicle to a training data center.
• metadata is transmitted along with the sensor data.
• Metadata may include the time of day, a timestamp, the location, the type of vehicle, vehicle control and/or operating parameters such as speed, acceleration, braking, whether autonomous driving was enabled, steering angle, odometry data, etc.
• Additional metadata includes the time since the last previous sensor data was transmitted, the vehicle type, weather conditions, road conditions, etc.
• the transmitted data is anonymized, for example, by removing unique identifiers of the vehicle.
• data from similar vehicle models is merged to prevent individual users and their use of their vehicles from being identified.
• the data is only transmitted in response to a trigger.
• an inaccurate prediction triggers the transmitting of image sensor and auxiliary sensor data for automatically collecting data to create a curated set of examples for improving the prediction of a deep learning network.
• a prediction performed at 405 to estimate the distance and direction of a vehicle using only image data is determined to be inaccurate by comparing the prediction to distance data from an emitting distance sensor.
• the image sensor data and related auxiliary data are transmitted and used to automatically generate training data.
• the trigger may be used to identify particular scenarios such as sharp curves, forks in the roads, lane merges, sudden stops, intersections, or another appropriate scenario where additional training data is helpful and may be difficult to gather.
• a trigger can be based on the sudden deactivation or disengagement of autonomous driving features.
• vehicle operating properties such as the change in speed or change in acceleration can form the basis of a trigger.
• a prediction with an accuracy that is less than a certain threshold triggers transmitting the sensor and related auxiliary data.
• a prediction may not have a Boolean correct or incorrect result and is instead evaluated by determining an accuracy value of the prediction.
• the sensor and related auxiliary data are captured over a period of time and the entire time series of data is transmitted together.
• the time period may be configured and/or be based on one or more factors such as the speed of the vehicle, the distance traveled, the change in speed, etc.
• the sampling rate of the captured sensor and/or related auxiliary data is configurable. For example, the sampling rate is increased at higher speeds, during sudden braking, during sudden acceleration, during hard steering, or another appropriate scenario when additional fidelity is needed.
• FIG. 5 is a diagram illustrating an example of capturing auxiliary sensor data for training a machine learning network.
• autonomous vehicle 501 is equipped with at least sensors 503 and 553 and captures sensor data used to measure object properties of neighboring vehicles 511, 521, and 561.
• the captured sensor data is captured and processed using a deep learning system such as deep learning system 100 of Figure 1 installed on autonomous vehicle 501.
• sensors 503 and 553 are additional sensors 103 of Figure 1.
• the data captured is the data related to vision data received at 203 of Figure 2 and/or part of the sensor data received at 401 of Figure 4.
• sensors 503 and 553 of autonomous vehicle 501 are emitting distance sensors such as radar, ultrasonic, and/or lidar sensors.
• Sensor 503 is a forward facing sensor and sensor 553 is a right-side facing sensor. Additional sensors, such as rear facing and left-side facing sensors (not shown) may be attached to autonomous vehicle 501.
• Axes 505 and 507 shown with long-dotted arrows, are reference axes of autonomous vehicle 501 and may be used as reference axes for data captured using sensor 503 and/or sensor 553. In the example shown, axes 505 and 507 are centered at sensor 503 and at the front of autonomous vehicle 501.
• an additional height axis (not shown) is used to track properties in three-dimensions.
• alternative axes may be utilized.
• the reference axis may be the center of autonomous vehicle 501.
• each sensor of sensors 503 and 553 may utilize its own reference axes and coordinate system. The data captured and analyzed using the respective local coordinate systems of sensors 503 and 553 may be converted into a local (or world) coordinate system of autonomous vehicle 501 so that the data captured from different sensors can be shared using the same frame of reference.
• field of views 509 and 559 of sensors 503 and 553, respectively are depicted by dotted arcs between dotted arrows.
• the depicted fields of views 509 and 559 show the overhead perspective of the regions measured by sensors 503 and 553, respectively.
• Properties of objects in field of view 509 may be captured by sensor 503 and properties of objects in field of view 559 may be captured by sensor 553.
• distance, direction, and/or velocity measurements of objects in field of view 509 are captured by sensor 503.
• sensor 503 captures the distance and direction of neighboring vehicles 511 and 521. Sensor 503 does not measure neighboring vehicle 561 since neighboring vehicle 561 is outside the region of field of view 509.
• the distance and direction of neighboring vehicle 561 is captured by sensor 553.
• objects not captured by one sensor may be captured by another sensor of a vehicle.
• autonomous vehicle 501 may be equipped with multiple surround sensors (not shown) that provide 360 degrees of visibility around the vehicle.
• sensors 503 and 553 capture distance and direction measurements.
• Distance vector 513 depicts the distance and direction of neighboring vehicle 511
• distance vector 523 depicts the distance and direction of neighboring vehicle 521
• distance vector 563 depicts the distance and direction of neighboring vehicle 561.
• the actual distance and direction values captured are a set of values corresponding to the exterior surface detected by sensors 503 and 553.
• the set of distances and directions measured for each neighboring vehicle are approximated by distance vectors 513, 523, and 563.
• sensors 503 and 553 detect a velocity vector (not shown) of objects in their respective fields of views 509 and 559.
• the distance and velocity vectors are three-dimensional vectors.
• the vectors include height (or altitude) components (not shown).
• bounding boxes approximate detected objects including detected neighboring vehicles 511, 521, and 561.
• the bounding boxes approximate the exterior of the detected objects.
• the bounding boxes are three-dimensional bounding boxes such as cuboids or another volumetric representation of the detected object.
• the bounding boxes are shown as rectangles around neighboring vehicles 511, 521, and 561.
• a distance and direction from autonomous vehicle 501 can be determined for each point on the edge (or surface) of a bounding box.
• distance vectors 513, 523, and 563 are related data to vision data captured in the same moment.
• the distance vectors 513, 523, and 563 are used to annotate distance and direction of neighboring vehicles 511, 521, and 561 identified in the corresponding vision data.
• distance vectors 513, 523, and 563 may be used as the ground truth for annotating a training image that includes neighboring vehicles 511, 521, and 561.
• the training image corresponding to the captured sensor data of Figure 5 utilizes data captured from sensors with overlapping fields of view and captured at matching times.
• a training image is image data captured from a forward facing camera that only captures neighboring vehicles 511 and 521 and not neighboring vehicle 561
• only neighboring vehicles 511 and 521 are identified in the training image and have their corresponding distance and directions annotated.
• a right-side image capturing neighboring vehicle 561 includes annotations for the distance and direction of only neighboring vehicle 561.
• annotated training images are transmitted to a training server for training a machine learning model to predict the annotated object properties.
• the captured sensor data of Figure 5 and corresponding vision data are transmitted to a training platform where they are analyzed and training images are selected and annotated.
• the captured data may be a time series of data and the time series is analyzed to associate the related data to objects identified in the vision data.
• FIG. 6 is a diagram illustrating an example of predicting object properties.
• analyzed vision data 601 represents the perspective of image data captured from a vision sensor, such as a forward-facing camera, of an autonomous vehicle.
• the vision sensor is one of vision sensors 101 of Figure 1.
• the vehicle's forward environment is captured and processed using a deep learning system such as deep learning system 100 of Figure 1.
• the process illustrated in Figure 6 is performed at 307, 309, and/or 311 of Figure 3 and/or at 401, 403, 405, 407, and/or 409 of Figure 4.
• analyzed vision data 601 captures the forward facing environment of an autonomous vehicle.
• Analyzed vision data 601 includes detected vehicle lane lines 603, 605, 607, and 609.
• the vehicle lane lines are identified using a deep learning system such as deep learning system 100 of Figure 1 trained to identify driving features.
• Analyzed vision data 601 also includes bounding boxes 611, 613, 615, 617, and 619 that correspond to detected objects.
• the detected objects represented by bounding boxes 611, 613, 615, 617, and 619 are identified by analyzing captured vision data. Using the captured vision data as input to a trained machine learning model, object properties such as distances and direction of the detected objects are predicted.
• velocity vectors are predicted.
• the detected objects of bounding boxes 611, 613, 615, 617, and 619 correspond to neighboring vehicles.
• Bounding boxes 611, 613, and 617 correspond to vehicles in the lane defined by vehicle lane lines 603 and 605.
• Bounding boxes 615 and 619 correspond to vehicles in the merging lane defined by vehicle lane lines 607 and 609.
• bounding boxes used to represent detected objects are three- dimensional bounding boxes (not shown).
• the object properties predicted for bounding boxes 611 are predicted for bounding boxes 611,
• FIG. 613, 615, 617, and 619 are predicted by applying a machine learning model trained using the processes of Figures 2-4.
• the object properties predicted may be captured using auxiliary sensors as depicted in the diagram of Figure 5.
• Figure 5 and Figure 6 depict different driving scenarios- Figure 5 depicts a different number of detected objects and in different positions compared to Figure 6-a trained machine learning model can accurately predict object properties for the objects detected in the scenario of Figure 6 when trained on sufficient training data.
• the distance and direction is predicted.
• the velocity is predicted.
• the predicted properties may be predicted in two or three-dimensions.
• the identified objects and corresponding properties can be used to implement autonomous driving features such as self-driving or driver-assisted operation of a vehicle.
• autonomous driving features such as self-driving or driver-assisted operation of a vehicle.
• a vehicle's steering and speed may be controlled to maintain the vehicle safely between two lane lines and at a safe distance from other objects.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Physics & Mathematics (AREA)
• Computer Vision & Pattern Recognition (AREA)
• Software Systems (AREA)
• Evolutionary Computation (AREA)
• Artificial Intelligence (AREA)
• Multimedia (AREA)
• Data Mining & Analysis (AREA)
• Medical Informatics (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• Health & Medical Sciences (AREA)
• Databases & Information Systems (AREA)
• General Health & Medical Sciences (AREA)
• Mathematical Physics (AREA)
• Life Sciences & Earth Sciences (AREA)
• Bioinformatics & Cheminformatics (AREA)
• Bioinformatics & Computational Biology (AREA)
• Evolutionary Biology (AREA)
• Traffic Control Systems (AREA)
• Image Analysis (AREA)
• Measuring And Recording Apparatus For Diagnosis (AREA)

Priority Applications (12)

Applications Claiming Priority (2)

Publications (1)

Family

ID=69784535

Family Applications (1)

Country Status (9)

Cited By (4)

Families Citing this family (78)

Citations (3)

Family Cites Families (836)