Risk maneuver assessment system for planning maneuvers with vehicle uncertainties

Abstract

Risk maneuver assessment system for planning maneuvers with uncertainties of a vehicle (100), comprising: a first control unit (34) with a processor (44) programmed to generate a perception of the vehicle's environment (100) and a behavior-decision model for the vehicle (100), including performing a calculation based on a sensor input to provide as output an action risk mapping and at least one target object tracking for different areas within the vehicle's environment (100); a sensor system (28) configured to provide the sensor input for the processor (44) to provide an area in the vicinity of the vehicle (100) for filtering target objects; one or more modules configured to map and track target objects via a processor in order to select a true candidate capture as the tracked target object from multiple candidate captures; one or more first modules configured to apply, via the processor, a Markov random field, MRF, algorithm to detect a current situation of the vehicle (100) in the environment and to predict an execution risk of a planned vehicle maneuver upon true detection of the dynamically tracked target; one or more second modules configured to apply mapping functions (610) via the processor to the captured environmental data in order to configure a machine learning model of the vehicle's decision behavior (100); one or more third-party modules configured, via the processor, to apply an adaptive threshold to cells of an occupancy grid configured to represent an area for tracking objects within the vehicle environment, and a second control unit with a processor configured to generate control commands according to the model of decision behavior and perception of the vehicle's environment (100) for planned vehicle maneuvers; one or more of the second and / or third modules, which are further configured to select at least one of the candidate captures that is within a radius around the target and / or a candidate capture that is closest to a first known mapped path, as the true capture; one or more of the second and / or third modules, which are further configured to select the candidate that indicates a position and velocity consistent with a target moving along a second known, mapped path, as the true capture, and to select as a false capture the candidate who indicates a position outside the second known path or whose speed does not match the target movement; and a fourth module configured to calculate, through a gating operation, a distance metric from the last position of a tracked target to a predicted position that is smaller than a threshold distance relating to one or more of the candidate captures.

Description

[0001] The present description generally refers to pre-decision risk planning and specifically to systems and procedures in a vehicle for generating more comprehensive representations of uncertainty indicators for use in pre-decision risk planning under conditions of reduced sensitive information.

[0002] Vehicle perception systems have been introduced into vehicles to allow them to perceive their surroundings and, in some cases, to enable autonomous or semi-autonomous navigation. Sensors that can be used in vehicle perception systems include radar, lidar, image sensors, and others.

[0003] Although significant progress has been made in vehicle perception systems in recent years, these systems still have room for improvement in several respects. Specifically, radars, particularly those used in the automotive industry, can provide data with considerable uncertainties regarding detection, speed, and position. This means that under conditions where radar scanning yields insufficient information and results, the necessary accuracy for further processing steps—such as target tracking and maneuver planning—cannot be achieved.

[0004] DE 10 2016 213 756 A1 describes a method for the safe operation of a motor vehicle, wherein a collision detection device determines whether a collision of the motor vehicle with a collision object is imminent and, upon detection of an imminent collision, a brake assistance device is activated to perform an emergency braking of the motor vehicle and a steering assistance device is activated to support a driver of the motor vehicle in an emergency evasive maneuver by applying a steering torque to a steering device of the motor vehicle as soon as the driver begins to steer.The steering assistance provided by the steering assistance device is made available before the emergency braking is carried out by the brake assistance device, whereby the brake assistance device automatically performs an emergency braking action on the motor vehicle as a result of its activation, after the steering assistance device has applied a steering torque to the steering device following its activation.

[0005] DE 10 2016 206 550 A1 describes a device and a method for determining the kinematics of a moving object with a course calculation filter for calculating an estimated direction of movement of the object based on a predicted position of the object and on the basis of the position of the object specified in radar measurement data of the object, and with a calculation unit for calculating Cartesian velocities of the object as a function of a measured radial object velocity and a measured object angle specified in the radar measurement data of the object, and as a function of the estimated direction of movement of the object calculated by the course calculation filter.

[0006] DE 10 2010 005 290 A1 describes a method for controlling a vehicle operating during a dynamic vehicle event, comprising monitoring a first input image, monitoring a first tracked object in the first input image in a first tracking cycle, monitoring a second input image, monitoring a second tracked object in the second input image in a second tracking cycle, and determining a distinctiveness measure that compares the first tracked object with the second tracked object. The distinctiveness measure estimates whether the first tracked object and the second tracked object represent a single tracked object in the vicinity of the vehicle.The procedure further includes assigning the first tracked object and the second tracked object based on the degree of difference, and using the assigned objects in a collision preparation system to control the operation of the vehicle.

[0007] US Patent 2009 / 0306881A1 describes a device and method that use a graphical model, such as a Markov random field model, to represent the principal directions of travel within an environment. The model has multiple nodes representing spatial positions within the environment. The principal direction for each node is determined probabilistically based on linear features detected in an image of the environment.

[0008] US 2014 / 0136045A1 describes a procedure encompassing controlling a vehicle according to a control strategy, detecting an object via sensors, predicting the future behavior of the object, and modifying the control strategy.

[0009] It can be considered a task to specify an improved system for the (as yet undecided) maneuver risk planning of vehicles prior to the decision by providing more comprehensive representations of uncertainty indicators from the results of radar scan acquisitions (or similar) of the vehicle, which leads to a reduction in the amount of data acquired, which can lead to incomplete representations of the uncertainties used to determine maneuver risk assessments for vehicle maneuvers.

[0010] The problem is solved by a system according to the invention as claimed in claim 1. Furthermore, a method, a vehicle and a planning system are described.

[0011] Systems and procedures are provided for an improved representation of uncertainty in radar detection for maneuver risk assessments.

[0012] A risk maneuver assessment system according to the invention for planning maneuvers involving vehicle uncertainties is described. The system according to the invention comprises: a first control unit with a processor programmed to generate a perception of the vehicle's environment and a behavioral decision model for the vehicle, including performing a calculation based on a sensor input to provide, as output, an action risk mapping and at least one target object tracking for various areas within the vehicle's environment; a sensor system configured to provide the sensor input to the processor to provide an area in the vehicle's environment for filtering target objects;one or more modules configured to map and track target objects using a processor in order to perform a candidate capture from multiple candidate captures to determine the true candidate capture as the tracked target object; one or more first modules configured to apply a Markov Random Field (MRF) algorithm from the processor to capture the current situation of the vehicle in the environment and predict the execution risk of a planned vehicle maneuver upon true capture of the dynamically tracked target; one or more second modules configured to apply mapping functions from the processor to captured environmental data in order to configure a machine learning model of the vehicle's decision-making behavior;one or more third modules configured to apply an adaptive threshold via the processor to cells of an occupancy grid configured to represent an area for tracking objects within the vehicle environment; and a second control unit with a processor configured to generate control commands in accordance with the model of the vehicle's decision-making behavior and perception of its environment for planned vehicle maneuvers.

[0013] The system according to the invention further comprises: the second and / or third module, which is configured to select, as the true capture, at least one candidate capture located within a radius around the target, and a candidate capture that is closest to a first known mapping path. The system further comprises: one or more of the second and / or third module(s), which is / are further configured to select, as the true capture, the candidate that indicates a position and velocity consistent with a target moving along a second known mapped path, and, as the false capture, the candidate that indicates a position outside the second known path or whose velocity does not correspond to that of the moving target.

[0014] The system according to the invention further comprises: a fourth module configured to calculate, by means of a gating operation, a distance metric from the last position of a tracked target object to a predicted position, which is less than a threshold distance relating to one or more of the candidate captures.The system further comprises: one or more fifth modules configured to apply the Markov Random Field (MRF) algorithm through the processor, which represents the tracked target object in one or more cells of an occupancy grid by: calculating an object measurement density for each tracked target object represented in the one or more cells of the occupancy grid; distributing the density over a window containing the cell set of the occupancy grid represented by the tracked target object; distributing the velocities over the window containing the same cell set of the occupancy grid represented by the tracked target object.

[0015] In one embodiment, the system further comprises: one or more first modules programmed to generate a Markov random field (MRF) for capturing the current situation.

[0016] In one embodiment, the system further comprises: one or more sixth modules configured to apply mapping functions from the processor to captured environmental data in order to configure a machine learning (ML) model of the vehicle's decision-making behavior through an action risk assessment model, which is trained using semi-supervised machine learning techniques through online and offline training to map the function to possible actions in order to determine a learned drivable route using risk factors.The system further includes: a seventh module configured to be executed by a processor in the offline training of the ML model, including: collecting labels, co-collecting occupancy velocity grids, extracting features from the occupancy grids, and applying at least Support Vector Machine (SVM) techniques to capture class patterns of candidate actions in order to determine the learned drivable path with risk factors.

[0017] In one embodiment, the system further comprises: an eighth module configured to apply, by the processor, an adaptive threshold to cells of an occupancy grid configured for representing the tracking area of ​​objects in the vehicle environment, including: a ninth module configured to calculate, by the processor, the probability that a candidate action is available for the target tracking object based on an adaptive threshold occupancy density and to select the candidate action with the highest probability of being available; and a tenth module configured to calculate, by the processor, a grouping of velocity clusters for a set of candidate actions in order to select the target tracking object that indicates a position consistent with a learned drivable path.The ML model is trained using reinforcement learning techniques with a dataset of past collected labels and sensor data of drivable paths, and the eight module is configured to select the candidate action that is likely to contribute to one of the drivable paths, with the sensor data including at least one of the following: radar, acoustic, lidar or image sensor data.

[0018] Furthermore, as an application of the risk maneuver assessment system, a vehicle is described that includes: a sensor detection probe comprising one or more from a set that includes: a radar, acoustic, lidar, and image sensor device; a risk maneuver assessment system for evaluating one or more uncertainties in planned maneuvers; and a variety of modules configured to generate, through a processor, a perception of the vehicle's environment and provide an output target for tracking various areas within the environment. The vehicle includes: the variety of modules including: one or more modules configured to map and track target objects through a processor in order to make a candidate detection from multiple candidate detections of a true candidate as the tracked target object;one or more modules configured to apply a Markov Random Field (MRF) algorithm via the processor to detect the current situation of the vehicle in the environment and to predict the execution risk of a planned vehicle maneuver when the dynamically tracked target is actually detected; one or more modules configured to apply mapping functions via the processor to detected environmental data in order to configure a machine learning model of the vehicle's decision-making behavior; one or more modules configured to apply an adaptive threshold via the processor to cells of an occupancy grid configured to represent areas of object tracking within the environment;and a control system with a processor configured to generate control commands in accordance with the model of the vehicle's decision-making behavior and perception of its environment for planned vehicle maneuvers.

[0019] The vehicle further comprises one or more modules programmed to generate a Markov Random Field (MRF) for capturing the current situation. The vehicle further comprises one or more modules configured to be selected as the true capture, including: a first module configured to select as the true capture the candidate capture located within a radius of the target; and a second module configured to select as the true capture the candidate capture closest to a first known mapped path.

[0020] The vehicle further comprises: one or more modules, which are further configured to select as the true capture, including: a third module configured to select the true capture, the candidate that indicates a position and velocity consistent with a target moving along a second known motion mapping path; and the third module configured to select a false capture, the candidate that indicates a position outside the second known path or whose velocity is not consistent with the target's motion.

[0021] The vehicle further comprises: one or more modules configured to be selectable as true captures, including: a fourth module configured to calculate, through a gating operation, a distance metric from the last position of a tracked target to a predicted position less than a threshold distance relating to one or more of the candidate captures.The one or more modules configured by the processor for the application of the Markov Random Field (MRF) algorithm, which represents the tracked target object in one or more cells of an occupancy grid; a fifth module is configured to calculate an object measurement density for each tracked target object represented in one or more cells of the occupancy grid; a sixth module is configured to distribute the density over a window containing the cell set of the occupancy grid represented by the tracked target object; a seventh module is configured to distribute the velocities over the window containing the same cell set of the occupancy grid represented by the tracked target object.

[0022] The vehicle further comprises that the one or more modules configured by the processor to apply mapping functions to captured environmental data in order to configure a machine learning (ML) model of the vehicle's decision-making behavior further comprises: an eighth module comprising an action risk assessment model, which is trained using semi-supervised machine learning techniques through online and offline training for mapping functions to possible actions in order to determine a learned drivable path with risk factors, wherein the eighth module in the offline training of the ML model comprises: collecting labels, co-collecting occupancy velocity grids, extracting features from the occupancy grids, and applying at least support vector machine (SVM) techniques to capture class patterns of candidate actions in order to determine the learned drivable path with risk factors.

[0023] The vehicle further comprises one or more modules for applying an adaptive threshold to cells of an occupancy grid configured to represent the area of ​​tracking objects within the vehicle environment, comprising: a ninth module configured to calculate, by means of an adaptive occupancy density threshold, the probability that a candidate action is available for the tracked target object based on the calculated density distribution, and to select the candidate action with the highest probability of being available; and a tenth module configured to calculate a velocity cluster grouping for a set of candidate actions in order to select the tracked target object that indicates a position consistent with a learned drivable path.The ML model is trained using reinforcement learning techniques, employing a dataset of previously collected labels and radar data of drivable routes, and the eight modules are configured to select the candidate action that is likely to contribute to one of the drivable routes.

[0024] Furthermore, a vehicle planning system is described. The system comprises: a sensor system configured to provide sensor input to the processor to define an area in the vehicle's environment for filtering target objects; and a non-temporary, computer-readable medium including: a first module configured by a processor to select as true capture the capture of the candidate that lies within a radius for the target; a second module configured by a processor to select as true capture the capture of the candidate that is closest to a first known mapped path;a third module configured to select, by means of a processor, the true capture, the candidate indicating a position and velocity consistent with a target moving along a second known, mapped path; and the third module configured to select a false capture, the candidate indicating a position outside the second known path or whose velocity does not match that of the moving target; a fourth module configured to calculate, by means of a processor, a distance metric from the last position of a tracked target object to a predicted position less than a threshold distance relating to one or more of the candidate captures;A fifth module is configured to calculate an object measurement density for each tracked target object represented in one or more cells of the occupancy grid using a processor; a sixth module is configured to distribute the density over a window containing the cell set of the occupancy grid represented by the tracked target object using a processor; a seventh module is configured to distribute the velocities over the window containing the same cell set of the occupancy grid represented by the tracked target object using a processor;an eighth module containing an action risk assessment model trained using semi-supervised machine learning techniques through online and offline training for mapping functions to candidate actions to determine a learned drivable route using risk factors, wherein the eighth module in the offline training of the ML model is configured to collect labels through a processor, collect occupancy velocity grids, extract features from the occupancy grids, and apply at least support vector machine (SVM) techniques to capture class patterns of the candidate actions to determine the learned drivable route using risk factors;a ninth module configured to use a processor to calculate the probability that a candidate action is available for the tracked target object based on the calculated density distribution using an adaptive threshold occupancy density, and to select the candidate action with the highest probability of being available;and a tenth module configured to compute velocity clustering for a set of candidate actions through a processor in order to select the targeted object that indicates a position consistent with a learned drivable path, wherein the ML model is trained using reinforcement learning techniques using a dataset of previously collected labels and radar data of drivable paths, and wherein the eighth module is configured to select the candidate action that is likely to contribute to one of the drivable paths.

[0025] The exemplary embodiments are described below in conjunction with the following figures, where similar reference numerals denote similar elements and where Fig. Figure 1 shows an example vehicle that includes a module for planning radar detection risk maneuvers, which more accurately represents the uncertainty indicators from the detected radar data in various embodiments; Fig. Figure 2 is a functional block diagram representing an autonomous driving system (ADS) associated with an autonomous vehicle according to various embodiments; Fig. Figure 3 is a block diagram that shows an example of a radar detection module for maneuver risk planning for use in a vehicle according to various embodiments; Fig. Figure 4 is a diagram illustrating by way of example the use of a pre-filter module with track cut-off to select the more probable true detection among several radar candidate detections received by radar sensors according to different embodiments; Fig. Figure 5 is a diagram illustrating the use of a Markov Random Field (MRF) module to generate a belief-state Markov model for tracking uncertainties in cross-traffic detection in accordance with various embodiments; Fig. Figure 6 is a diagram illustrating the use of an example action risk assessment module that uses path plan data through semi-supervised modeling and machine learning (ML) training to select the more likely true candidate action among multiple candidate action captures in order to predict drivable paths in accordance with different embodiments; Fig. Figure 7 is a diagram illustrating the use of an exemplary object extraction for object tracking using adaptive threshold occupancy grid density cells in accordance with various embodiments; and Fig. Figure 8 is a process flow diagram that illustrates an example process for maneuver risk assessment of uncertainties for predicting drivable paths based on multiple radar candidate detections in accordance with different embodiments.

[0026] The term "module" here refers to any hardware, software, firmware, electronic control component, processing logic and / or processor device, individually or in any combination, including but not limited to: application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), electronic circuit, processor (common, dedicated or grouped) and memory executing one or more software or firmware programs, combined logic circuit and / or other suitable components providing the described functionality.

[0027] Embodiments of the present description can be described here in the form of functional and / or logical block components and various processing steps. It should be noted that such block components can be implemented by any number of hardware, software, and / or firmware components configured to perform the specified functions. Thus, an embodiment of the present description may employ various integrated circuit components, such as memory elements, digital signal processing elements, logic elements, lookup tables, or the like, which, under the control of one or more microprocessors or other control devices, can perform a variety of functions.Furthermore, those who are familiar with the art will understand that embodiments of the present description can be practiced in conjunction with any number of systems and that the systems described here are merely exemplary embodiments of the present description.

[0028] Autonomous vehicles operating in complex, dynamic environments require methods that can generalize to unpredictable situations and react promptly to achieve a level of reliability and safe response, even in complex urban environments. Informed decision-making requires accurate perception. However, current computer vision systems have yet to achieve error rates acceptable for autonomous navigation. Combining decision-making, control, and perception with machine learning techniques and complex planning and decision-making methods, as described here, can be a viable option.

[0029] For the sake of brevity, conventional techniques of signal processing, data transmission, signaling, control, machine learning, radar, lidar, image analysis, and other functional aspects of the systems (and the individual operating components of the systems) are not described in detail here. Furthermore, the connecting lines shown in the various figures are intended to represent exemplary functional relationships or physical couplings between the different elements. It should be noted that many alternative or additional functional relationships or physical connections may exist in an embodiment of the present description.

[0030] The subject matter described herein comprises devices, systems, techniques, and articles for the operation of a vehicle's maneuver control and planning systems. The described devices, systems, techniques, and articles are connected to a vehicle's sensor system and a controller for receiving inputs from one or more sensor devices of the sensor system, including uncertainty levels, for use in determining, planning, predicting, and / or executing vehicle maneuvers in real time, in the near future, or in the future. For this purpose, the controller may employ at least one adaptive algorithm that uses uncertainty indicators based on limited available sensor data and / or on a priori sensor data.

[0031] The sensors can include a combination of sensors with different operating modes to acquire a variety of sensor data. For example, the sensors can include one or more cameras as well as radar- or laser-based sensors (e.g., lidar sensors). That is to say, while the description in detail details radar-based scan acquisitions, it is considered that the description should not be so restrictive and should include a variety of different sensor devices used individually or in combination, such as acoustic, lidar, image, and IR sensors.

[0032] During operation, the sensor system can collect sensor data that is received by a processor in the control unit. The processor can be programmed to transform the sensor data into a perceptual model (i.e., belief) with uncertainties about the vehicle and / or its environment in order to either make a decision or plan a maneuver prior to making a decision. For example, the processor can determine the location of surrounding vehicles relative to the vehicle in question, predict the path of surrounding vehicles, determine and track the current vehicle path, detect lane markings, locate pedestrians and cyclists and predict their movements, and, with representations of uncertainties of the decisions to be made, predict future vehicle maneuvers and more.

[0033] In some embodiments, the processor generates an occupancy grid with a plurality of cells that collectively represent the vehicle's perceived environment. The processor computes at least one perception datum for the different cells within the occupancy grid. The perception datum represents a perceived element of the vehicle's environment. The processor generates a representation of an uncertainty (i.e., an uncertainty factor) for the different cells, where the uncertainty representation may indicate the processor's uncertainty about the perception and future maneuver associated with the cell, as a means of a more comprehensive representation and taking into account factors of this uncertainty.The perception data and uncertainty factors can be calculated from the received sensor output for the vehicle sensors using one or more estimation applications such as Bayesian, Markov or other statistical algorithms.

[0034] The perception (i.e., the representation of uncertainty) and the individual uncertainty factors contained in the network cells can be continuously updated during vehicle operation. Additionally, the processor determines the situational relevance of the various cells within the grid. This relevance can be determined in several ways.

[0035] The devices, systems, and procedures presented here can utilize the results from other sensors present in a vehicle (e.g., cameras, other radar devices) to provide a more complete representation of the uncertainty, in order to determine the risk indicators before deciding on the maneuver, thus avoiding the problem of reduced information due to the representation during the planning of the vehicle maneuver.

[0036] In various exemplary embodiments, the decision indicators can be combined with other pre-decision indicators to achieve better planning under uncertainty.

[0037] In various exemplary embodiments, the described apparatus, systems, and methods can be combined and the results of back-end processing methods, which provide data products for perception fusion, can be improved via a representation to achieve more accurate maneuver planning under uncertainty.

[0038] The apparatuses, systems, and procedures disclosed here employ a more complete representation of uncertainty to determine risk indicators for pre-decision maneuvers, thereby avoiding the problem of reduced information due to the representation. Furthermore, the pre-decision indicators can be combined with other multiple approaches and / or other pre-decision indicators and planning objectives to achieve improved planning under uncertainty.

[0039] In various embodiments, the present description offers apparatus, systems, and methods, using an occupancy-speed-belief-state density grid and a belief-state Markov model, to track the uncertainty about the detection and awareness of cross traffic.

[0040] In various embodiments, this description presents apparatus, systems and methods that perform supervised, unsupervised and estimated modeling, which learns the risk of performing certain maneuvers, such as merging into traffic at an intersection with and without extraction of conventional object trajectories, and predicts it through improved uncertainty representations in order to easily merge sensor data with the outputs of other sensors (e.g. cameras, lidars) of the vehicle.

[0041] In various embodiments, the present description offers apparatus, systems and methods for vehicle maneuvers of: merging into traffic, planning of merging maneuvers under high uncertainty perception, more robust handling of uncertainty, of complex (multimode) vehicle presence distribution estimations and of transmitting information of higher complexity directly to discriminating pre-decision data products.

[0042] In various embodiments, the present description offers apparatus, systems and methods for vehicle maneuvers: from several approaches of varying complexity, from amplification learning of maneuver risks, from occupancy-speed grid-Markov model tracking, from the use of additional sensors, whether existing or standalone, and from integration with existing traditional tracking and fusion systems.

[0043] In various forms, the present description offers apparatus, systems and procedures for maneuver risk planning over specific regions of interest and only for the current state of belief by looking into the state of belief space in order to determine only those states of belief that are reachable from the current state.

[0044] Instead of keeping the perception modules and the risk maneuver planning modules separate, an alternative framework is described here for training specific parts of the perception module in order to integrate subtasks from the risk maneuver planning module.

[0045] Fig. Figure 1 shows an example vehicle 100, which includes a planning module for cross-traffic tracking and maneuver risk estimation to generate a more comprehensive representation of the uncertainty indicators 302 (hereinafter referred to as the “maneuver risk planning module”) from the acquired data of a radar system (and / or other sensors). As shown in Fig. As shown in Figure 1, the vehicle 100 generally consists of a chassis 12, a body 14, the front wheels 16, and the rear wheels 18. The superstructure 14 is arranged on the chassis 12 and essentially encloses components of the vehicle 100. The superstructure 14 and the chassis 12 can together form a frame. The wheels 16-18 are each rotatably coupled to the chassis 12 near a corner of the superstructure 14.

[0046] In various embodiments, the vehicle 100 can be an autonomous vehicle or a semi-autonomous vehicle. An autonomous vehicle 100, for example, is a vehicle that is automatically controlled to transport passengers from one place to another. In the illustrated embodiment, the vehicle 100 is depicted as a passenger car, but other vehicle types such as motorcycles, trucks, sport utility vehicles (SUVs), recreational vehicles (RVs), watercraft, aircraft, etc., can also be used.

[0047] As shown, the vehicle 100 generally comprises a drive system 20, a transmission system 22, a steering system 24, a braking system 26, a sensor system 28, an actuator system 30, at least one data storage device 32, at least one control unit 34, and a communication system 36. The drive system 20 can, in various embodiments, comprise an internal combustion engine, an electric machine such as a traction motor, and / or a fuel cell drive. The transmission system 22 is configured to transmit power from the drive system 20 to the vehicle wheels 16 and 18 according to selectable speed ratios. Depending on the embodiment, the transmission system 22 can comprise a stepped automatic transmission, a continuously variable transmission, or another suitable transmission.

[0048] The braking system 26 is configured to supply braking torque to the vehicle wheels 16 and 18. The braking system 26 can, in various embodiments, include friction brakes, cable brakes, a regenerative braking system such as an electric motor, and / or other suitable braking systems.

[0049] The steering system 24 influences the position of the vehicle wheels 16 and 18, respectively. Although a steering wheel 25 is shown for illustration purposes, the steering system 24 may not include a steering wheel in some embodiments considered within the scope of this description.

[0050] The sensor system 28 comprises one or more sensor devices 40a-42n that detect observable conditions of the external environment and / or the internal environment of the vehicle 100 (such as the condition of one or more occupants) and generate corresponding sensor data. The sensor devices 40a-42n may include, among others, radars (e.g., long-range, medium-range, and short-range), lidar, global positioning systems, optical cameras (e.g., forward-facing, 360°, rear-facing, side-facing, stereo, etc.), thermal imaging cameras (e.g., infrared cameras), ultrasonic sensors, odometry sensors (e.g., encoders), and / or other sensors that can be used in conjunction with systems and methods in accordance with the present subject matter.

[0051] The actuator system 30 comprises one or more actuator devices 40a-42n that control, but are not limited to, one or more vehicle features such as the drive system 20, the transmission system 22, the steering system 24, and the braking system 26. In various embodiments, the vehicle 100 may also be operated in Fig. 1. Interior and / or exterior features of the vehicle shown, such as various doors, a trunk and cabin features such as air, music, lighting, touchscreen display components (e.g. in conjunction with navigation systems) and similar items.

[0052] The data storage device 32 stores data for the automatic control of the vehicle 100. In various embodiments, the data storage device 32 stores defined maps of the navigable environment. In these embodiments, the defined maps can be predefined by and obtained from a remote system. The defined maps can, for example, be compiled by the remote system and transmitted to the vehicle 100 (wirelessly and / or via cable) and stored in the data storage device 32. Route information can also be stored in the data storage device 32 – that is, a series of road segments (geographically linked to one or more of the defined maps) that together define a route the user can take to travel from a starting point (e.g., the user's current location) to a destination.As can be estimated, the data storage device 32 can be part of the control unit 34, separate from the control unit 34, or part of the control unit 34 and part of a separate system.

[0053] The control unit 34 contains at least one processor 44 and a computer-readable storage device or medium 46. The processor 44 can be any custom or commercially available processor, a central processing unit (CPU), a graphics processing unit (GPU), an application-specific integrated circuit (ASIC) (e.g., a custom ASIC implementing a neural network), a field-programmable gate array (FPGA), an auxiliary processor among several processors assigned to the control unit 34, a semiconductor-based microprocessor (in the form of a microchip or chipset), any combination thereof, or, more generally, any device for executing instructions. The computer-readable storage device or medium 46 can, for example, include volatile and non-volatile storage in a read-only memory (ROM), a random-access memory (RAM), and a keep-alive memory (KAM).KAM is a persistent or non-volatile memory that can be used to store various operating variables when the processor 44 is switched off. The computer-readable storage device or computer-readable data carrier 46 can be implemented with any number of known storage devices such as PROMs (programmable read-only memory), EPROMs (electrical PROMs), EEPROMs (electrically erasable PROMs), flash memory, or other electrical, magnetic, optical, or combined storage devices capable of storing data, some of which represent executable instructions used by the control unit 34 to control the vehicle 100. In various implementation variants, the control unit 34 is configured to implement a mapping system, as discussed in detail below.

[0054] The instructions can contain one or more separate programs, each containing an ordered list of executable instructions for implementing logical functions. When executed by the processor 44, the instructions receive and process signals (e.g., sensor data) from the sensor system 28, perform logic, calculations, procedures, and / or algorithms for the automatic control of the vehicle 100's components, and generate control signals that are transmitted to the actuator system 30 to automatically control the vehicle 100's components based on the logic, calculations, procedures, and / or algorithms. Although in Fig. 1 where only one control unit 34 is shown, embodiments of the vehicle 100 may contain any number of control units 34 which communicate via any suitable communication medium or combination of communication media and which cooperate to process the sensor signals, perform logic, calculations, procedures and / or algorithms and generate control signals for automatic control of features of the vehicle 100.

[0055] The communication system 36 is configured to wirelessly transmit information to and from other units 48, such as other vehicles (“V2V” communication), infrastructure (“V2I” communication), networks (“V2N” communication), pedestrians (“V2P” communication), remote transportation systems, and / or user devices. In one exemplary embodiment, the communication system 36 is a wireless communication system configured to communicate via a wireless local area network (WLAN) using the IEEE 802.11 standard or via cellular data communication. However, additional or alternative communication methods, such as a dedicated short-range communication (DSRC) channel, are also considered in this briefing.DSRC channels refer to one-way or two-way short- to medium-range wireless communication channels specifically designed for automotive applications and a corresponding set of protocols and standards.

[0056] Depending on various embodiments, the controller 34 can be an autonomous drive system (ADS) 70 according to Fig. 2. This means that suitable software and / or hardware components of the control unit 34 (e.g. processor 44 and computer-readable memory 46) can be used to implement an autonomous driving system 70 that is used in conjunction with the vehicle 100.

[0057] In various embodiments, the instructions of the autonomous driving system 70 can be organized according to function or system. For example, the autonomous driving system 70 can, as in Fig. Figure 2 shows a perception system 74, a positioning system 76, a path planning system 78, and a vehicle control system 80. As can be estimated, the instructions can be organized in various embodiments in any number of systems (e.g., combined, further subdivided, etc.), since the description is not limited to the examples presented.

[0058] The perception system 74 synthesizes and processes the acquired sensor data in various configurations and predicts the presence, location, classification, and / or path of objects and features in the vehicle 100's environment. In various embodiments, the perception system 74 can acquire information from multiple sensors (e.g., sensor system 28), including, but not limited to, cameras, lidars, radars, and / or any number of other sensor types. In various embodiments, the radar acquisitions can be fully or partially integrated into the perception system 74.

[0059] The positioning system 76 processes sensor data along with other data to determine the position (e.g., a local position relative to a map, a precise position relative to a lane of a road, a vehicle direction, etc.) of the vehicle 100 relative to its environment. As one can imagine, a variety of techniques can be used to achieve this localization, including, for example, simultaneous localization and mapping (SLAM), particle filters, Kalman filters, Bayesian filters, and the like.

[0060] The path planning system 78 processes the sensor data together with other data to determine a path that the vehicle 100 is to follow. The vehicle control unit 80 generates control signals to control the vehicle 100 according to the determined route.

[0061] In various embodiments, the control unit 34 implements machine learning techniques to support the functionality of the control unit 34, such as feature detection / classification, obstacle mitigation, route crossing, mapping, sensor integration, ground truth determination, and similar functions.

[0062] In various embodiments, the positioning system 76 is configured to determine where the vehicle 100 is located or positioned within the grid (i.e., the occupancy grid), and the dynamic object detection systems determine where moving objects are located relative to the vehicle 100 within the grid (not shown). The sensor inputs of sensors 40a-40n can be processed by the maneuver risk planning module 302 to perform these determinations. Furthermore, in some embodiments, the vehicle positioning system 76 and / or the path planning system 78 communicate with the other units to determine the relative positions of the vehicle 100 and surrounding vehicles, pedestrians, cyclists, and other dynamic objects.

[0063] More specifically, the sensor input of the Risk Maneuver Planning Module 302 can consist of radar and / or laser (Lidar) detections from one or more of the 40a-40n sensor devices. The Maneuver Risk Planning Module 302 can filter and determine which of the detections are dynamic objects (moving objects that are actually on the road).

[0064] The Maneuver Risk Planning Module 302 can process this information and generate a Markov Random Field (MRF) (i.e., Markov mesh, undirected graphical model, etc.) to represent the dependencies within it. Using this information and an enhancement training process, the Maneuver Risk Planning Module 302 can determine (i.e., predict) the risk associated with initiating (i.e., executing) a particular maneuver (e.g., a right turn into cross traffic).

[0065] From this prediction function, the maneuver risk planning module 302 can determine the degree to which individual cells influence the risk prediction output. In some embodiments, the maneuver risk planning module 302 can identify which of the sensors 40a-40n have the greatest influence on the maneuver risk prediction, and these sensors 40a-40n can correlate with specific cells. Those cells that have a greater influence on the risk prediction are identified by the maneuver risk planning module 302 as more relevant than the others.

[0066] Fig. Figure 3 is a high-level block diagram depicting an example maneuver risk planning module 302 for providing data for assessing action risks and extracting objects for use within behavioral and perception systems of a vehicle according to an embodiment. The example maneuver risk planning module 302 is configured to apply one or more tracking and estimation modules to a plurality of radar scan acquisitions from a radar 305 (with n scans (ΔT)) for target acquisition, with pre-filtering of the acquired data by a pre-filtering module 310 to configure an occupancy velocity-belief state density grid.

[0067] In various exemplary configurations, the Maneuver Risk Planning module determines the relevance uncertainty indicators (also called uncertainty factors) that contain relevance uncertainty indicators for the different areas. This involves processing sensor data to: recognize the current state of the vehicle and accordingly precalculate the uncertainty indicators for assessing the risk of executing a specific vehicle maneuver; determine, prior to decision-making in the planning phase, the degree of influence that a particular uncertainty indicator has on the forecast for the different areas; and calculate the maneuver risk indicator for the different areas according to the determined degree of influence, including calculating higher levels of maneuver risk indicators for areas with higher degrees of influence.The control command of the AV perception module 335 also includes the generation of the control command for the maneuver depending on the uncertainty indicator and the risk relevance factor of the maneuver.

[0068] The Maneuver Risk Planning Module 302 is programmed to generate a Markov Random Field (MRF) from a Markov Random Field Module 315 to detect the current situation and apply a belief-state Markov model to specific uncertainty indicators about the target objects (i.e., cross traffic) from radar scan data received over a specified period. The multiple modules in the example Maneuver Risk Planning Module 302 include the Markov Random Field (MRF) Module 315 for detecting current situations, generating occupancy speed grid data, and populating cells in the occupancy speed grid. The Action Risk Assessment Module 320 receives data in occupancy speed cells and generates uncertainty indicators for actions, which are then mapped to the cells for action risk mapping and preliminary decision-making.The object extraction module 325 generates object tracking based on data from cells that are populated in the occupancy velocity grid for common perception areas from the input of the radar sensor 305, in order to provide object tracking data according to the relevance uncertainty indicators for the pre-decisions for different areas within perception.

[0069] The example pre-filter module 310 also includes a pre-filter module 310 configured to select one or more mapping and gating modules to determine which of several radar target acquisitions is a likely acquisition. The example maneuver risk planning module 302 is further configured to output the likely acquisition determined by the pre-filter module 310. The radar acquisition for the maneuver risk planning module 302 can be accessed on the control unit 34 of Fig. 1 or implemented on a separate control unit or on a combination of control units in various embodiments.

[0070] In Fig. 4. The pre-filter module 310 performs map filter steps from the acquired road data and the road speed data from the radar scan acquisitions of radar 305 (with n scans (ΔT)) to configure an occupancy speed-belief state density raster. In the map matching module 410, the scan acquisitions are mapped between the sensor domain and the map domain.

[0071] The example card matching module 410 is according to Fig. 4 is configured to select the more likely true detection of target 421 from among several false / true radar candidates by choosing a candidate that is closest to a known target path and less than certain maximum distance thresholds from the target path. The selection module (road speed filter 418) can select the map matching module 410 to choose the more likely true detection of target 421 if a map of the area near the vehicle is available and / or alternative true / false detection methods are not available.

[0072] The example map matching module 410 is configured to use the position of a road on an area map to determine valid target positions when an area map is available. The example map matching module 410 is also configured to infer a valid target position from camera data, route plan data, and map data (e.g., lane, curb detection, target loops). The example map matching module 410 is configured to select the candidate detection that is closest to a known target path (e.g., the center of the road) and less than a maximum threshold distance (determined by a calculated function of angle and distance from the host to the target point) from the target path (i.e., the center line of the road). The map can be a detailed, pre-existing map or a coarse map section derived from the detection results of an imaging device, such as a radar.a camera, was derived.

[0073] After receiving the road data and an estimation of the speed of the vehicle and the moving targets, map fitting can be applied to the target position. This involves measuring the distance between the projected points (i.e., the true moving targets), P1, P2, and P3, within a radius defined by a circle 412 from each of the points P1 to P3. Target points (in this case, P1 to P3) within a map fitting area are then determined by the pre-filter module 310 to distinguish between false detections 415 and true detections 420, based on the expected speed, the direction of the moving target, and assumptions about its movement around the center of the road. In other words, the pre-filter module calculates the apparent static range of the target and its measured speed on the road to determine whether it is a false detection 415 or a true detection 420.

[0074] The example false and true detections on map module 308 are configured, at discrete times, to: (a) calculate the apparent static distance rate (srr) of each target; (b) calculate the road speed of each target, assuming the target is moving on the road; and (c) filter out (e.g., reject) candidate detections that do not match a target moving on the road. In this example, target 416 is not positioned on the road at this time, and the associated candidate detection is filtered out. In the example, target 421 is positioned in a lane on the road, its direction of travel matches the direction of travel of the lane, and its speed matches the speed limit for the lane. Consequently, the candidate detection associated with target 421 is not filtered out in this case.

[0075] Following this determination of the predicted position by the distance clock module 425 (or charge clock) of the received measurement data, a predicted position is calculated based on the last position. Example module 425 is configured to select the more probable true detection from among several uncertain candidate detections in circle 412 by choosing the candidate closest to a predicted target position as the true radar detection. Map matching module 410 can select route timing module 425 if radar tracks already exist for a target and / or alternative tracking and estimation methods are unavailable.

[0076] Example module 425 is configured to (a) calculate a distance metric to each true candidate detection (i.e., P1, P2, P3) for each existing radar track; (b) predict a nearest position to a target position using a prediction filter (e.g., a Kalman filter); and (c) select the true candidate detections (i.e., target 421) that are closest to the predicted nearest position and less than a threshold distance from it. The distance metric may include, but is not limited to, the Euclidean distance or the Mahalanobis distance; the latter can be calculated using the predicted position covariance obtained from a Kalman filter.

[0077] In one example, the track clock module 425 can select a candidate that is closest to a predicted target position 401 for radar acquisition by calculating a distance metric from the last position 403 of a tracked target to each candidate acquisition for each existing radar track; applying a Kalman filter to predict the next position of the tracked target and converting the predicted next position into the measurement plane (e.g., for linear KF, y). - = HFx̂); Calculation of the squared Mahalanobis distance dk2=(y−)TS−1y− where S=R+HP−HT for the detection of candidates k ∈ {1,2, ...} (can alternatively use the Euclidean distance); and, for each existing radar track 411, the gating of candidate detection, knowing that dk2 is chi-squared distributed to choose a threshold T percent P (e.g. 95%), and the associated capture k with the orbit, if dk2 <T. In the event that multiple uncertain detections are located within gate 405, the detection with the shortest Mahalanobis distance is selected.

[0078] Example module 425 is configured to access a prediction filter (e.g., a Kalman filter) to obtain a predicted position 401 for a destination based on the last position 403 for that destination. Using the predicted position 401, example route clock module 425 is configured to identify a gate region 405 (e.g., create a gate) that is within a threshold distance of the nearest predicted position 401. Example module 425 is configured to select the probable true capture by choosing the nearest true candidate capture (e.g., candidate capture for destination 421) that lies within gate region 405 as the probable true capture.

[0079] Fig. Figure 5 shows a diagram of the Markov Random Field (MRF) module for capturing the current situation and applying a belief-state Markov model to specific uncertainty indicators about the target objects (i.e., cross traffic) from radar scan data received over a predetermined period, according to one embodiment. Module MRF 315 (i.e., Markov grid, non-directional graphical model, etc.) is used to represent the dependencies for the uncertainty indicators for pre-decision planning. Using this information and an enhancement training process, the maneuver risk planning module 302 can determine (i.e., predict) the risk associated with initiating (i.e., executing) a specific maneuver (e.g., a left or right turn into cross traffic).

[0080] In Fig. 5 receives the Markov random field (MRF) module 315 filtered static and dynamic targets that have been clustered into dynamic objects by clustering and module 505 (from Fig. 5) for capturing current situations, generating occupancy rate raster data, and populating cells of the occupancy rate raster. The MRK module 315 in a measurement raster aggregation module 510 populates a measurement raster with dynamic objects from the clustering step. The measurement raster aggregation module 510 goes through a feedback loop of processing steps to incrementally populate each available cell of the raster until the entire pipeline of each received dynamic object is exhausted, or the raster is considered sufficiently populated (i.e., after a predetermined time period), or until completion.The cyclic processing steps for each dynamic object include calculating the object measurement density 515, distributing the density across the 2D cell windows 520, distributing the velocities across the same 2D cell windows 525, and adding the calculated scatter densities and velocities of each object to the raster 530. This cycle is then repeated until the raster is complete or deemed sufficiently complete. After the dynamic objects have been inserted into the raster, the densities are summed by a computational function using module 540. Subsequently, the velocities across the cells of the cell window are averaged by a computational function or weighted by factors using a weighted average velocity module 545 for density-velocity measurement raster representations.

[0081] In various exemplary embodiments, a Bayesian update 550 is provided for each dynamic object for exemplary density and velocity modeling on the measurement aggregation grid. The Bayesian update 550 includes the update rate based on the Bayesian occupancy weight 560 from the input of an update occupancy probability module 555 calculation. In some embodiments, the MRF module 315 calculates the uncertainty values ​​for the cells 552 using one or more Bayesian algorithms. The calculations are used to quantify, for various cells 552, the expected error (i.e., the information gain), which is calculated as true occupancy (Ø ∈ {0,1}), minus an estimate (p) squared, multiplied by the probability with respect to the occupancy.In this context, occupancy grid algorithms are used to compute approximate post-estimations for these random variables. In other words, the expected prediction error (i.e., the uncertainty factor) calculated by Module 555 for each grid cell by updating the grid occupancy probabilities can be calculated according to the following equation (1): E[(∅−p)2]=∑∅∈(0,1)(∅−p)2P(∅)=(0−p)2(1−p)+(1−p)2p=p(1−p) where ø represents the actual occupancy, p the estimate, and P the probability. Additionally, a Bayesian update 550 can be performed for a given cell 552 according to the following equation (2): pPost=(1−a)(n−k)akp(1−a)(n−k)akp+(1−b)(n−k)bk(1−p) where n is the number of observations at a given cell, k is the number of detections, a is the detections (P), b is the false alarm (P), and p is the occupancy (P).

[0082] Accordingly, the perception module 335 (of Fig. 3) For a given cell 552, calculate the backspaces according to equation (2). Additionally, the perception module 335 can calculate the expected future uncertainty for cells 552 for use in risk assessments according to the following equation (3): E[RMSE]=p(1−p)(ab+(1−a)(1−b))n

[0083] The perception module 335 can thus create a heuristic model that can be used to compensate for uncertainties and to adaptively control the sensors 40a-42n. For a specific cell 552 within the grid, the perception module 335 determines how much the uncertainty increases or decreases when one or more sensors 40a-42n are used in the corresponding physical space within the environment defined by the occupancy grid. In some embodiments, the adaptive sensor control system 34 relies on this information (uncertainty reduction in the cells 552) when it generates sensor control commands to the sensors 42a-42n.

[0084] Fig. Figure 6 shows a schematic of a module for assessing the risk of action for processing data from cells of the occupancy rate grid according to one embodiment. Fig. 6. The action risk assessment module receives 320 data points in cells of the occupancy rate and applies mapping functions with uncertainty indicators for candidate actions for use in action risk mapping and pre-decision-making.

[0085] The example of action risk assessment module 320 in Fig. The 6 is configured to select the more probable true candidate action by applying a mapping function 610 based on the captured occupancy rate data 605 from the occupancy rate grid 600. Each captured occupancy rate data 605 from each cell, when applying a mapping function 610, is capable of generating multiple action candidates 620. Using a trained machine learning (ML) model 615 with the risk vectors 625 for each of the multiple action candidates 620, the pre-decision maneuver behavior can be modeled.

[0086] For example, the action candidates 620 in the trained ML model 615 can include straight, right, and left predecision maneuvers. The ML model 615 is configured to predict drivable paths, determine which of the risk vectors contributes to the candidate actions in predicting the drivable paths, and select candidate actions that contribute to the drivable paths as the more probable true candidate actions.

[0087] Furthermore, offline semi-monitoring learning can be modeled through an example of offline reinforcement learning, in which the offline training module 640 contributes to action risk mapping using historical, stored data or cloud data from labels consisting of a video of the location in question and the corresponding sensor data. The reinforcement learning example offline training module 640 includes a trained machine learning model trained to predict future drivable paths through traffic using candidate action captures.The example reinforcement learning (discriminator) offline module 640 is configured to collect labels 645 for determining maneuvers such as "Is it safe to turn left?"; to collect occupancy grids (MRF) 650 for applying mapping functions 610 for offline candidate action generation; to extract features from grid 655 into the trained ML model to predict drivable paths; and to apply support machine vectors (SVM) and Gaussian elimination 660 to capture class patterns of candidate actions in order to determine which of the candidate actions contribute to the drivable paths, and to select candidate actions that contribute to the drivable paths as the more probable true candidate actions to send to the ML model 615, thereby enabling the application of learned mapping functions in the ML model 615.By applying both labels and risk vector indicators, the behavior module can collect 630 behavioral labels such as "Is it safe to do behavior B?"...etc. in real time.

[0088] In various exemplary embodiments, a collected, labeled dataset of historical radar data, such as speed data and drivable paths, is used. This dataset can be generated from data about the paths previously traveled by the vehicle. The historical sensor data can be used for training with enhanced lemming techniques, specifically the ML model 615, to predict drivable paths based on candidate action detections. Once the ML model 615 has been trained to predict drivable paths based on candidate action detections, real-time sensor data can be applied to the ML model.

[0089] Fig. Figure 7 shows a schematic of the object extraction module for generating object traces for preliminary decision-making by the risk maneuver planning module, according to one embodiment. The object extraction module 325 (of Fig. 7) Generates object tracks based on data from cells populated in the occupancy velocity grid 710 for perception areas, in order to provide object track data for publication according to the relevance uncertainty indicators represented for different areas within a perception grid. The occupancy velocity grid 710 is configured with an adaptive threshold of the occupancy density cells 715 by adaptively configuring the inputs of the threshold 718 to ensure that the uncertainty indicators are effectively integrated into the occupancy velocity grid 710. This means that locations with a high probability use a physically based bottom-up clustering 720 and where velocity cluster centers 725 can be associated with the object tracks 730, regions with high uncertainty can be distinguished from those with low uncertainty.The process includes feedback on the previously detected speed clusters 727 for efficiency. The resulting object tracks 730 are published in object representations for machine learning models of vehicle perception.

[0090] Fig. Figure 8 is a process flow diagram that presents an example process for maneuver risk assessments with uncertainties to predict drivable routes based on multiple radar candidate detections in accordance with various embodiments.

[0091] In the flowchart of Fig. In Task 810, sensor data is generated by a number of radar scans over a time period and prepared for processing before reception. It is considered that a variety of sensor devices, including near- and far-range, image, light, acoustic, and IR sensor devices, can be used or incorporated to generate the acquired data sets, and the description is not limited to radar data.

[0092] Next, task 820 involves processing steps for mapping and tracking dynamic targets. The pre-filter module (i.e., 310 of Fig. 4) performs map filtering steps on captured data from road data and road speed data from the radar scan captures (with n scans (ΔT)) to configure an occupancy speed belief state density speeder representing regions of interest above the vehicle. The sample map matching module selects the most probable true target capture from several false / true radar candidates by choosing a candidate that is closest to a known target path and less than a maximum distance threshold from the target path. The selection module can choose the map matching module to select the more probable true target capture if a map of the area near the vehicle is available and / or alternating true / false capture methods are not available.The sample map matching module uses the position of a road on an area map to determine valid target positions, if an area map is available. The sample map matching module is also configured to infer a valid target position from camera data, railway plan data, and map data (e.g., lane, curb detection, target rails). The sample map matching module is configured to select the candidate detection that is closest to a known target path (e.g., the center of the road) and less than a maximum threshold distance (determined by a calculated function of the angle and distance from the host to the target point) from the target path (i.e., the center line of the road). The map can be a detailed, pre-existing map or a coarse map section derived from the detection results of an imaging device, such as a camera.

[0093] In task 830, the MRK algorithm is used to generate a raster-based dynamic model in order to determine uncertainty indicators about the target objects (i.e., cross traffic). The MRF (i.e., 315 of Fig. 5) Determines (i.e., predicts) the risk associated with initiating (i.e., executing) a specific maneuver (e.g., a left or right turn into cross traffic). The processing steps include the Markov Random Field (MRF) module, which receives filtered static and dynamic targets to capture current situations, generate occupancy speed grid data, and populate cells of the occupancy speed grid; the application of a measurement grid aggregation module to populate a measurement grid with dynamic objects from the clustering step; and a feedback loop of processing steps to incrementally populate each available cell of the grid until the entire pipeline of each received dynamic object is exhausted, the grid is deemed sufficiently populated (i.e., within a prescribed timeframe), or completion.The cyclical processing steps for each dynamic object include calculating the object measurement density, distributing the density across 2D cell windows, distributing the velocities across the same 2D cell windows, and adding the calculated scatter densities and velocities of each object to the raster. This cycle is then repeated until the raster is full or deemed sufficiently complete. After the dynamic objects are added to the raster, the densities are summed, and the velocities are averaged across the cells or weighted with factors to enable the display of the density-velocity measurement rasters. Additionally, a Bayesian update is provided for the sample density and velocity modeling, which includes updating the velocities based on the Bayesian occupancy weighting.

[0094] Task 840 involves conducting an action risk assessment using semi-supervised machine learning (ML) model training with offline training. The sample action risk assessment module selects the more probable true candidate actions by applying a mapping function based on the captured occupancy velocity data and then applying a trained machine learning (ML) model with risk vectors for each of the multiple candidate actions, allowing the behavior prior to the maneuver decision to be modeled. For example, the candidate actions might incorporate straight, right, and left pre-decision maneuvers into the trained ML model.The ML model predicts drivable paths, determines which of the risk vectors contributes to the candidate actions in order to predict the drivable paths, and selects the candidate action that contributes to the drivable paths as the more likely true candidate action for behavior modeling 850.

[0095] Furthermore, tasks related to offline semi-supervised reinforcement learning (840) are modeled by an example of offline reinforcement learning that contributes to action risk mapping using historical, stored data or cloud data from labels that form a coherent video of the location in question and corresponding sensor data. This is an example.An offline reinforcement learning-trained machine learning (ML) model predicts future drivable paths through traffic by using candidate action captures by collecting labels to determine maneuvers such as "is it safe to turn left?"; co-collecting occupancy rasters (MRF) to apply mapping functions to generate candidate actions offline; extracting features from rasters to train the ML model to predict drivable paths; and finally applying support machine vectors (SVM) and Gaussian processing to capture class patterns of candidate actions to determine which of the candidate actions contribute to the drivable paths for behavior modeling 850.

[0096] Task 845 describes the processing steps for object extraction with an adaptive threshold of the grid density for object-based representations. The object extraction module (i.e., 325 of Fig.7) Performs processing steps to generate object tracking traces based on data from cells populated in the occupancy velocity grid for perceptual areas from the acquired data, in order to provide object tracking data for publication according to the relevance uncertainty indicators represented for different areas within a perceptual grid. In other words, a certain influence of the different areas is determined by the densities of the occupancy grid. The occupancy velocity grid is configured with an adaptive threshold of the occupancy density cells by receiving threshold inputs to adaptively configure the threshold occupancy cell densities to ensure that the uncertainty indicators are effectively integrated into the occupancy velocity grid.This means that locations with a high probability of using physically based bottom-up clustering, and where velocity cluster centers can be associated with object trajectories, regions with high uncertainty can be distinguished from those with low uncertainty. The process includes feedback of the previously identified velocity clusters for efficiency. The object trajectories resulting from the processing steps are published in Object Representations for ML Models of Vehicle Perception 855.

[0097] The devices, systems, methods, techniques, and articles described here can be applied to measurement systems other than radar systems (e.g., lidar, acoustic, or imaging systems). The devices, systems, methods, techniques, and articles described here can also be applied to speed sensors such as laser- or light-based speed sensors.

Claims

[1] Risk maneuver assessment system for planning maneuvers with uncertainties of a vehicle (100), comprising: a first control unit (34) with a processor (44) programmed to generate a perception of the vehicle's environment (100) and a behavior-decision model for the vehicle (100), including performing a calculation based on a sensor input to provide as output an action risk mapping and at least one target object tracking for different areas within the vehicle's environment (100); a sensor system (28) configured to provide the sensor input for the processor (44) to provide an area in the vicinity of the vehicle (100) for filtering target objects; one or more modules configured to map and track target objects via a processor in order to select a true candidate capture as the tracked target object from multiple candidate captures; one or more first modules configured to apply, via the processor, a Markov random field, MRF, algorithm to detect a current situation of the vehicle (100) in the environment and to predict an execution risk of a planned vehicle maneuver upon true detection of the dynamically tracked target; one or more second modules configured to apply mapping functions (610) via the processor to the captured environmental data in order to configure a machine learning model of the vehicle's decision behavior (100); one or more third-party modules configured, via the processor, to apply an adaptive threshold to cells of an occupancy grid configured to represent an area for tracking objects within the vehicle environment, and a second control unit with a processor configured to generate control commands according to the model of decision behavior and perception of the vehicle's environment (100) for planned vehicle maneuvers; one or more of the second and / or third modules, which are further configured to select at least one of the candidate captures that is within a radius around the target and / or a candidate capture that is closest to a first known mapped path, as the true capture; one or more of the second and / or third modules, which are further configured to select the candidate that indicates a position and velocity consistent with a target moving along a second known, mapped path, as the true capture, and to select as a false capture the candidate who indicates a position outside the second known path or whose speed does not match the target movement; and a fourth module configured to calculate, through a gating operation, a distance metric from the last position of a tracked target to a predicted position that is smaller than a threshold distance relating to one or more of the candidate captures. [2] System according to claim 1, further comprising: that one or more first modules are programmed to generate a Markov random field, MRF, in order to recognize the current situation. [3] System according to claim 1, further comprising: one or more fifth modules, configured via the processor, apply the Markov Random Field (MRF) algorithm, which represents the tracked target object in one or more cells of an occupancy grid, by: Calculating an object measurement density for each tracked target object represented in one or more cells of the occupancy grid; Distributing the density over a window encompassing the cell set of the occupancy grid represented by the tracked target object; and Distributing velocities across the window that encompasses the same set of cells in the occupancy grid represented by the tracked target object. [4] System according to claim 3, further comprising: one or more sixth modules configured to apply mapping functions (610) via the processor to captured environmental data to configure a machine learning (ML) model of the vehicle's decision behavior (100) through an action risk assessment model trained using semi-supervised machine learning techniques through online and offline training for the mapping function (610) to the candidate actions in order to determine a learned drivable path with the risk factors. [5] System according to claim 4, further comprising: a seventh module, configured to run via a processor during offline training of the ML model, comprising: Collecting labels (645), Collection of occupancy speed grids, Extracting features from the occupancy grids, and Applying at least support vector machines (SVM) techniques to detect class patterns of candidate actions in order to determine the learned drivable path using risk factors. [6] System according to claim 5, further comprising an eighth module configured to apply, via the processor, an adaptive threshold to cells of an occupancy grid configured to represent the area of ​​tracking objects within the vehicle environment, comprising: a ninth module that is configured, via the processor, to calculate, based on an adaptive threshold occupancy density, the probability that a candidate action is available for the tracked target object based on the calculated density distribution, and selects the candidate action that has the highest probability, to be available; and a tenth module that is configured, via the processor, to calculate a clustering for velocity clusters for a set of candidate actions in order to select the tracked target object that indicates a position consistent with a learned drivable path.

Images