Stable radar tracking speed initialization using multiple hypotheses

Abstract

This document describes an object tracker that uses multiple hypotheses to perform stable velocity initialization for radar tracking, including when only sparse radar point clouds are available. In the case of only a single point per scan, the tracker creates multiple hypotheses for the object's direction and velocity. A least squares function can be applied to each hypothesis to derive each respective initial velocity, which is tracked using a Kalman filter during a hypothesis tracking time period. As hypotheses are initialized and tracked over each hypothesis tracking time period, their tracking error scores are computed. Based on the hypothesis' tracking error score, hypotheses with low evidence are discarded during the hypothesis tracking time period. When the hypothesis tracking time period ends, the hypotheses with high evidence initialize the tracked velocity. Parallel hypothesis evaluation enables the tracker to quickly and accurately initialize velocity by selecting only the best hypothesis, which can enable safer driving.

Description

Background Technology

[0001] Perception systems for vehicles (e.g., advanced safety or autonomous driving systems) can rely on the output of radar trackers. Each track represents a point cloud of radar echoes detected over multiple frames, which can be grouped to represent individual objects. The initial velocity for tracking can be derived based on the variance of the position of points measured over multiple frames using point cloud detection. It is important to initialize the tracking velocity measurements as accurately as possible; otherwise, tracking splits may occur. For example, if the object's velocity is incorrectly initialized, the reported object position may deviate from its true position. For more distant objects, a sparse point cloud (e.g., a single point) may be all the information available for tracking generation, making velocity initialization difficult. To address this, filters or advanced algorithms can be used, including linear least squares-based algorithms such as iterative least squares or normalized estimation of the squared error. However, the results using these techniques can be too unstable to be used with vehicle control, which expects accurate tracking to ensure safety. Furthermore, even if an object can be detected at a great distance from a single point, erroneous decisions made using incorrect assumptions about the object's position, orientation, or speed can propagate downstream to the user tracking it, potentially leading to unsafe or uncomfortable driving. Existing processes fail to provide consistent or stable results, contributing to these problems and reducing safety due to the inability to accurately track objects. Summary of the Invention

[0002] This document describes techniques and systems for stable radar tracking velocity initialization using multiple hypotheses. In one example, the method includes obtaining point cloud sensor data from a radar system by an object tracker, the point cloud sensor data indicating radar echoes reflected from objects in the environment; and establishing tracking of objects in the environment using the point cloud sensor data. The method includes initializing an object velocity measurement by creating multiple hypotheses about the predicted movement of the object; determining an initial velocity and a first associated level of evidence supporting that initial velocity for each of the multiple hypotheses; generating a fused hypothesis by combining the multiple hypotheses, the initial velocity of the fused hypothesis and the first associated level of evidence being based on a set of the initial velocities and first associated levels of evidence for the multiple hypotheses; including the fused hypothesis among the multiple hypotheses; and, in response to including the fused hypothesis among the multiple hypotheses, selecting a first optimal hypothesis based on the first associated level of evidence for the multiple hypotheses for initializing the object velocity measurement.

[0003] In some examples, the above method further includes: updating the initial velocity of each of the multiple hypotheses over time; determining a second associated level of evidence supporting the updated initial velocity for each of the multiple hypotheses; and eliminating any hypothesis from the multiple hypotheses that has a value of the second associated level of evidence that does not meet the evidence threshold. Furthermore, in response to determining that two or more hypotheses remain after eliminating any hypothesis with a value of the second associated level of evidence that does not meet the evidence threshold, the above method further includes: updating the predicted movement of the object for each of the remaining multiple hypotheses; and selecting a second best hypothesis to replace the previously selected first best hypothesis based on the second associated level of evidence of the remaining multiple hypotheses. In response to determining that only two hypotheses remain after eliminating any hypothesis with a value of the second associated level of evidence that does not meet the evidence threshold, the above method further includes: selecting a third best hypothesis to replace the previously selected first best hypothesis or second best hypothesis based on the second associated level of evidence of the remaining two hypotheses; and outputting tracking of the object to the vehicle system, the tracking including a velocity parameter initialized to the initial velocity of the third best hypothesis.

[0004] By implementing these and other examples conceived by this disclosure, stable radar tracking velocity initialization using multiple assumptions can be achieved to initialize velocity measurements more accurately than using other radar tracking techniques, even when only a single point return is available for the tracked object per frame. This summary introduces a simplified concept of stable radar tracking velocity initialization using multiple assumptions, for example, in vehicles equipped with radar tracking to support driving (e.g., trucks, cars), and other examples of object tracking, as further explained in the detailed description and accompanying drawings. This summary is not intended to identify essential features of the claimed subject matter, nor is it intended to define the scope of the claimed subject matter. Attached Figure Description

[0005] This document describes in detail the initialization of a stable radar tracking velocity using various assumptions, with reference to the accompanying drawings. The drawings may use the same numbers to refer to similar features and components, and hyphenated numbers to specify variations of these similar features and components. The drawings are organized as follows:

[0006] Figure 1 An example environment for stable radar tracking velocity initialization using multiple assumptions is shown;

[0007] Figure 2 An example vehicle is shown, comprising a system configured to perform stable radar tracking speed initialization using multiple assumptions;

[0008] Figure 3-1 A flowchart is shown as an example procedure for initializing a stable radar tracking velocity using multiple assumptions;

[0009] Figure 3-2 and Figure 3-3 It shows in Figure 3-1 An example scenario in the context of the main vehicle initializing radar tracking speed measurement; and

[0010] Figure 4 A graph showing the initial velocity measurement over a period of time using stable radar tracking velocity initialization with various assumptions is presented. Detailed Implementation

[0011] introduction

[0012] Radar systems measure the radial velocity of points on an object; this is called the range-rate. Multiple points (e.g., a radar point cloud) or multiple frames (e.g., a radar scan) are needed to extract the object's velocity profile. Radar systems typically provide a sparse point cloud of some objects (e.g., including only a few points) during each frame or scan, or even just a single point at a considerable distance to detect objects, known as long-range objects. Because there are only a limited number of points or a single point in the radar point cloud, it is difficult to accurately initialize the tracked velocity. The number of radar points in a single scan may be insufficient to obtain a stable solution. In other words, the velocity calculation may be unstable or inconsistent with the actual movement of the object.

[0013] A typical process for calculating velocity involves solving a least-squares problem using the rate of change of distance and azimuth angle calculated for different points within a radar point cloud. The positional changes of points are calculated across two or more consecutive radar scans. However, the instability of velocity results resulting from these existing techniques fails to provide accurate and predictable stable results from one frame to the next. For example, when multiple reflection points from an object are close to each other in each scan, the reflection points may actually contain almost linearly related information. However, if the reflection points are far apart in each scan, the positional changes of these points in consecutive scans can produce corrupted velocities because, for simplicity, the range of objects also captured by positional changes is often ignored. That is, positional changes between two points seem impossible given the size constraints of other objects on the road (especially vehicles). Existing algorithms and techniques attempt to address this problem by checking the consistency of positional information changes from radar points in consecutive cycles, trying to reduce the weight of inconsistent information in the final assumptions about the initial object motion and orientation.

[0014] This document describes an object tracker that performs stable velocity initialization for radar tracking using multiple hypotheses, including when only a sparse radar point cloud is available. In the case of a single point per scan, the tracker creates multiple hypotheses for the predicted movement of the object. A least-squares function can be applied to each hypothesis to derive each corresponding initial velocity, and these hypotheses are tracked using a Kalman filter during the hypothesis tracking period. While initializing and tracking the hypotheses during the hypothesis tracking period, their tracking error scores are calculated. Hypotheses with low evidence (e.g., high error scores) are discarded during the hypothesis tracking period. At the end of the hypothesis tracking period, a single hypothesis with high evidence (e.g., low error scores) is used to initialize the tracked velocity. Parallel hypothesis evaluation enables the tracker to initialize the tracking velocity quickly and accurately by selecting only the best hypothesis, which can lead to safer driving.

[0015] Example Environment

[0016] Figure 1 An example environment 100 is shown, initialized for, for example, a stable radar tracking speed of a vehicle 102 using various assumptions. Although shown as a passenger vehicle, the vehicle 102 could represent other types of motorized vehicles (e.g., cars, motorcycles, buses, tractors, semi-trailers), non-motorized vehicles (e.g., bicycles), rail vehicles (e.g., trains), water vehicles (e.g., boats), aircraft (e.g., airplanes), or spacecraft (e.g., satellites), etc. The depicted environment 100 includes the vehicle 102 traveling on a road. The vehicle 102 is equipped with a sensing system 104 for detecting objects 106 (or other similar objects) present on or near the road, which could affect how or whether the vehicle 102 can continue to travel.

[0017] The perception system 104 is used to relay information about objects detected in environment 100 (such as object 106 in the travel path of vehicle 102) to other systems of vehicle 102. A region of interest associated with the perception system 104 at least partially surrounds vehicle 102. This region is referred to as field of view 108 (also known as instrument field of view) when monitored by the perception system 104. The perception system 104 operates based on input from point cloud sensor data, which can be obtained from a single type of sensor or various different types of sensors. For ease of description, the point cloud sensor data obtained by the perception system 104 is primarily described as radar data obtained from the radar system of vehicle 102. The perception system 104 can be mounted on, integrated into, or incorporated into any part of vehicle 102, such as the front, rear, top, bottom, or sides of vehicle 102, bumpers, side mirrors, headlights and / or taillights, or any other internal or external location of vehicle 102 where object detection using point cloud sensor data is desired. Careful selection or repositioning of the components of the sensing system 104 and / or the radar system to which the sensing system 104 is connected can further give the field of view 108 a specific shape or size.

[0018] The perception system 104 includes a processor 110 and a radar interface 112. The processor 110 executes software and / or firmware that configures the perception system 104 to perform various functions to facilitate object tracking. For example, information output from the processor 110 may take the form of tracks 116; each track 116 points to a different object detected in the field of view 108. Tracks 116 may include information fields that include bounding box dimensions, size and position measurements, classification, and other data characterizing vehicles, pedestrians, traffic signs, or other objects present in the area monitored by the perception system 104. The processor 110 is configured to generate tracks 116 based on point cloud sensor data 114 received by the processor 110 from the radar interface 112.

[0019] Radar interface 112 may include a combination of hardware and software executed thereon or on processor 110. Radar interface 112 operatively connects processor 110 of perception system 104 to the output of a radar-based sensor source (including the radar system of vehicle 102 and / or an external radar source). Radar interface 112 can use vehicle-to-vehicle communication to obtain point cloud sensor data from the radar systems of other vehicles, for example, to improve the size or resolution of field of view 108. Radar interface 112 may provide point cloud sensor data in compressed or uncompressed form or in any other format suitable for object tracking. A communication channel is shared between radar interface 112 and the radar system. The communication channel may include an application programming interface (API) or other functions executed by processor 110. Radar interface 112 is an example sensor interface. Perception system 104 may also use another interface to another type of sensor (e.g., lidar, camera) to obtain additional point cloud sensor data of different sensor types.

[0020] The object tracker 118, including initializer 120, is an example component of perception system 104 that can be at least partially implemented in executable code, which, when executed, configures processor 110 to generate tracks 116 of objects in field of view 108. Tracks 116 are output from object tracker 118 to other systems of vehicle 102 that rely on tracks 116 to gain situational awareness of potential obstacles. When initializer 120 is executed, it configures processor 110 to initialize the corresponding velocity of each of the tracks 116 in a stable manner using a variety of assumptions.

[0021] Dependent on initializer 120, object tracker 118 is configured to perform stable velocity initialization of tracking 116 using multiple assumptions, including when only sparse radar point cloud is available. In the case of a single point per scan, initializer 120 creates multiple assumptions about the object's orientation and velocity. For example, object tracker 118 obtains point cloud sensor data 114 from radar interface 112, which indicates radar echoes reflected from objects in environment 100. Object tracker 118 performs object tracking techniques. Using point cloud sensor data 114, tracking of object 106 is added to tracking 116. However, instead of using other techniques to initialize the velocity measurement of object 106, initializer 120 is configured to initialize the velocity measurement more accurately and stably. Initializer 120 is configured to create at least two assumptions about the predicted movement of object 106. For each of the at least two assumptions, an initial velocity and a first associated level of evidence supporting that initial velocity are determined. Initializer 120 then generates fused hypotheses by combining at least two hypotheses. More specifically, the fused hypotheses comprise the initial velocity and the first associated evidence level of the fused hypotheses based on the initial velocity of the at least two hypotheses and a set of first associated evidence levels. Initializer 120 includes the fused hypotheses in the at least two hypotheses. Initializer 120 may apply a least-squares function to each hypothesis to derive each corresponding initial velocity.

[0022] In response to including the fused hypothesis among at least two hypotheses, initializer 120 is configured to select a single optimal hypothesis for initializing the velocity measurement of object 106 based on a first associated level of evidence for the at least two hypotheses. For example, initializer 120 updates the hypothesis using a Kalman filter during the hypothesis tracking period. During the hypothesis tracking period, each hypothesis is initialized and tracked by initializer 120, which includes initializer 120 calculating their tracking error scores. Hypotheses with low evidence (e.g., high error scores) may be discarded during the hypothesis tracking period. When the hypothesis tracking period ends, the single hypothesis with high evidence (e.g., low error scores) initializes the tracked velocity. This parallel hypothesis evaluation performed by initializer 120 in this manner enables object tracker 118 to quickly initialize the tracked velocity by selecting only the optimal hypothesis, ultimately enabling safer driving because tracking 116 used by other systems of vehicle 102 for control or safety functions includes more accurate information, but is as fast or nearly as fast as initialization in other ways.

[0023] Example vehicle configuration

[0024] Figure 2An example vehicle 102-1 is shown, comprising a system 104-1 configured to perform stable radar tracking speed initialization using multiple assumptions. Vehicle 102-1 is an example of vehicle 102.

[0025] The vehicle 102-1 includes a sensing system 104-1, which is an example of the sensing system 104 shown in more detail. The vehicle 102-1 also includes a vehicle-based system 210, which is operatively and / or communicatively coupled to the sensing system 104-1 via a link 202. The link 202 can be one or more wired and / or wireless links, including vehicle-based network communications for interconnecting components of the vehicle 102-1. In some examples, the link 202 is a vehicle communication bus.

[0026] The vehicle-based system 210 uses vehicle data (including object tracking data provided by the perception system 104-1 on link 202) to perform vehicle-based functions, which may include, among other functions, functions for vehicle control. The vehicle-based system 210 may include any conceivable device, apparatus, component, module, part, subsystem, routine, circuit, processor, controller, etc., that uses radar data to represent actions taken by the vehicle 102-1. As some non-limiting examples, the vehicle-based system 210 may include systems for autonomous control, systems for safety, systems for positioning, systems for vehicle-to-vehicle communication, systems for serving as occupant interfaces, and systems for serving as multi-sensor trackers. Upon receiving object tracking data (e.g., tracking 116), the functions provided by the vehicle-based system 210 use portions of the object tracking data (including velocity measurements of objects detected in the field of view 108) to configure the vehicle 102-1 for safe driving without colliding with the detected objects.

[0027] Tracking 116 is an example of object tracking data output on link 202 to vehicle-based system 210. One of the tracks in 116 may include information about the movement of object 106, such as speed, position, etc., enabling vehicle-based system 210 to control or assist braking, steering, and / or acceleration of vehicle 102-1 to avoid collision with object 106. Systems for autonomous control can use tracking 116 received via link 202 to drive vehicle 102-1 safely on a road, autonomously or semi-autonomously. Systems serving as occupant interfaces can use information in tracking 116 to allow operators or passengers to have situational awareness to make driving decisions or to provide operator input to controllers to provide more buffers to avoid objects. Systems for vehicle-to-vehicle communication can use tracking 116, or the information contained therein, to provide other vehicles with tracking 116, allowing operators, passengers, or controllers of other vehicles to also avoid tracked objects or to have confidence that vehicle 102-1 knows of their presence based on receiving tracking 116. By improving the situational awareness of vehicle 102-1 and other vehicles in the environment 100, vehicle 102-1 can be driven more safely under manual, autonomous, or semi-autonomous control.

[0028] The perception system 104-1 includes a processor 110-1 (example of processor 110), a radar system 206, and a computer-readable medium (CRM) 208. The radar system 206 may include any number of radar devices, antennas, and other components to provide point cloud sensor data 114 covering the field of view 108. The radar system 206 may include radar chips, antennas, or antenna arrays such as multiple-input multiple-output (MIMO). The radar system 206 may include various transmitter / receiver elements, timing / control elements, and analog-to-digital converters. As already mentioned, although described primarily in the context of radar, the perception system 104-1 may employ other sensor systems to perform point cloud-based object tracking.

[0029] Some examples of processor 110-1 include controllers, control circuitry, microprocessors, chips, systems, system-on-a-chip, devices, processing units, digital signal processing units, graphics processing units, and central processing units. Processor 110-1 may include multiple processors, one or more cores, embedded memory storing software or firmware, caches, or any other computer element that enables processor 110-1 to execute machine-readable instructions for generating trace 116.

[0030] Machine-readable instructions executed by processor 110-1 may be stored in CRM 208. CRM 208 may also be used to store data managed by processor 110-1 during instruction execution. In some examples, CRM 208 and processor 110-1 are a single component, such as a system-on-a-chip including CRM 208 configured as dedicated memory for processor 110-1. In some examples, access to CRM 208 is shared by other components of sensing system 104-1 (e.g., radar system 206) connected to CRM 208. Processor 110-1 obtains instructions from CRM 208; execution of the instructions configures processor 110-1 to perform object tracking operations (such as radar-based object tracking), which results in the transmission of tracking 116 via link 202 to vehicle-based system 210 and other components of vehicle 102-1.

[0031] In this example, CRM 208 includes instructions for configuring processor 110-1 to provide radar interface 112-1, which is an example of radar interface 112. Furthermore, CRM 208 includes instructions for executing object tracker 118-1 (including initializer 120-1), which are respectively... Figure 1 Examples of object tracker 118 and initializer 120 in the example.

[0032] In operation, object tracker 118-1 is configured to acquire point cloud sensor data 114 generated by radar system 206 from radar interface 112-1. Point cloud sensor data 114 transmits information about radar echoes detected from objects in environment 100. Object tracker 118-1 is configured to process point cloud sensor data 114 to establish tracks 116; each track 116 corresponds to a specific object in environment 100. When object tracker 118-1 generates a track 116, the track 116 is reported to vehicle-based system 210. Upon establishing each track 116, object tracker 118-1 invokes initializer 120-1 to initialize the velocity measurement of each track before that track is introduced into track 116.

[0033] To initialize a velocity measurement in track 116, initializer 120-1 generates and maintains hypothesis 212, which is partially based on solutions to a least linear squares (LLS) problem. Solving the LLS problem is simplified by using a weighted least squares principle. Hypothesis 212 initially includes at least two hypotheses corresponding to solutions to different LLS problems, and another hypothesis derived by fusing the solutions to all the different LLS problems into one. For example, hypothesis 212 can be derived from solving seven different LLS problems to potentially capture seven different potential radar point distributions from point cloud sensor data 114. The seven LLS problems produce seven solutions, which correspond to the seven hypotheses in hypothesis 212. An eighth hypothesis is added to hypothesis 212, and, as described in more detail below, the eighth hypothesis is formed by fusing the seven solutions. High accuracy is then achieved by tracking hypothesis 212 for a short period to select the optimal hypothesis.

[0034] After initializer 120-1 establishes hypothesis 212, the velocity measurement of track 116 is not immediately initialized. Instead, all hypotheses 212 are tracked using a motion model (e.g., a simple constant motion model) for several frames or scans (e.g., fewer than or equal to eight scans) of radar system 206. The number of frames scanned while tracking hypothesis 212 can be adjusted or regulated for different modes or characteristics of radar system 206, sensing system 104-1, and / or vehicle 102-1. During each scan of the hypotheses tracking time period, a tracking error score is calculated and assigned to each hypothesis in hypothesis 212.

[0035] The tracking error score is calculated based on the cumulative error between each hypothesis in Hypothesis 212 and the velocity measurement associated with the best hypothesis. The tracking error score is used by initializer 120-1 to select the best hypothesis for initializing one of the velocity measurements in tracking 116. Unlike other hypothesis tracking techniques that consider the association between each hypothesis and all measurements at each scan, the complexity of velocity initialization is reduced. Instead of considering every association at each scan, initializer 120-1 is configured to consider only a single best association at each scan. The tracking error score enables initializer 120-1 to arrive at a single best hypothesis more quickly by eliminating every unlikely hypothesis from Hypothesis 212 after each scan of the hypothesis tracking period until a single hypothesis is obtained.

[0036] After the initializer 120-1 determines the speed measurement for the new track, the new track is added to track 116. Track 116 can be updated with measurements for the current tracking reporting period. By using the speed measurement from the initializer 120-1, track 116 can be output with high accuracy and low latency. This enables the vehicle-based system 210, which receives the track on link 202, to perform the functions of vehicle 102-1 with increased safety and higher accuracy.

[0037] The process of initializing a stable radar tracking velocity using multiple assumptions

[0038] Figure 3-1 A flowchart of an example process 300 for initializing a stable radar tracking velocity using multiple assumptions is shown. For ease of description, process 300 is primarily described in the context of execution by sensing systems 104 and 104-1. The operations (also referred to as steps) of process 300 are numbered; however, this numbering does not necessarily imply a specific order of operations. The steps of process 300 can be compared with... Figure 3-1 The diagram shows different ways to rearrange, skip, repeat, or execute specific methods. Figure 3-2 and Figure 3-3 It shows in Figure 3-1 Example scenarios 330-1 and 330-2 are given in the context of the primary vehicle initializing radar tracking speed measurements. In each example, it is assumed that after four consecutive scans or frames of radar system 206, only a single point from point cloud sensor data 114 is associated with each detectable object. Those skilled in the art can readily generalize this example and, for example, apply process 300 to different scenarios where more than one point in point cloud sensor data 114 is associated with each detectable object by calculating, for example, the average position of all points associated with the detectable object within point cloud sensor data 114.

[0039] Assume an initialization time period (e.g., steps 302 to 308).

[0040] In operation, perception system 104 acquires point cloud sensor data 114 from radar system 206, which indicates radar echoes reflected from objects in environment 100. Using point cloud sensor data 114, perception system 104 establishes tracking of object 106 in the environment, including initializing velocity measurements of object 106. Perception system 104 may execute initializer 120 while executing object tracker 118. When initializer 120 is invoked by object tracker 118, initializer 120 execution process 300 initializes velocity measurements of object 106.

[0041] At point 302, hypotheses are created for at least two different distributions of the point cloud sensor data. For example, initializer 120 creates multiple hypotheses for the predicted movement of object 106. Two types of information can be used to solve for the initial velocity, as defined by Equation 1 for defining the rate of change of distance and azimuth information, and Equation 2 for defining the change of position:

[0042]

[0043] ΔT i v x =x i -x i-1 ,ΔT i v y =y i -y i-1 ,

[0044] In each of these equations 1 and 2, the subscript i indicates each different scan, and a constant rate is assumed within multiple scans (e.g., four scans) called the scan time interval.

[0045] consider Figure 3-2 The tracking 116 generated for scene 330-1 points to four different objects in the field of view 108 of the vehicle 102. Each of the tracking 116 includes the bounding box size of the different object. The bounding boxes are based on information inferred from single-point observations in the point cloud sensor data 114 over a scanning period (in this case, four scans). For example, a bounding box 332-1 is derived for the first object based on consecutive points 334-1, and another bounding box 332-2 is derived for the second object using consecutive points 334-2. For the third and fourth objects in the field of view 108, a third bounding box 332-3 can be derived from consecutive points 334-3, and a fourth bounding box 332-4 can be derived from consecutive points 334-4.

[0046] In this scenario, when determining the velocity measurement of each object within an object based on consecutive, single points observed over a scanning period, various factors need to be considered. For example, one of the tracks 116 corresponds to bounding box 332-1; changes in the longitudinal position of this track (e.g., as indicated by the X-axis) may not be particularly useful for estimating direction or velocity, but changes in the lateral position (e.g., as indicated by the Y-axis) may be useful. The opposite may be true for a different track 116; one of the tracks 116 corresponds to bounding box 332-2, and changes in the lateral position of this track may not be useful for determining velocity, but changes in the longitudinal position may be useful for such an estimation. For a third object associated with bounding box 332-3, changes in position in both directions (e.g., seemingly simultaneous changes in lateral and longitudinal positions) may be useful for deriving an initial velocity estimate for the third track in track 116. Conversely, for a fourth object associated with bounding box 332-4, changes in position in both directions are not useful for deriving an initial velocity estimate for the fourth track in track 116.

[0047] This indicates that if the angular separation θ between consecutive points used to derive a particular bounding box... i Too small (e.g., angular separation θ) i If the angle separation is less than the separation threshold, then the continuous points of the bounding box may be unreliable for determining the velocity. Similarly, in some cases, even if the angle separation θ is less than the separation threshold... i Large enough, continuous points are not useful for estimating speed. The object's speed (v) x v y It can be derived from continuous points based on Equation 1, where v x It is the longitudinal velocity of the object, v y It is the horizontal speed of the object. This is the rate of change of radial distance to the object, which compensates for the principal velocity. With the angular separation θ... i Approaching zero, the solution speed (v) x v y It's very difficult. Even the angular separation θ between consecutive points... i When the size is large enough, inconsistent positional changes at certain points (e.g., when some points cross the first edge of an object, while other consecutive points cross a second edge orthogonal to the first edge) cause a velocity (v) to be affected. x v y It is also unreliable.

[0048] When initializer 120 is executed by processor 110 at 302, initializer 120 of object tracker 118 can generate the following seven hypotheses, referred to as Hypothesis 1 to Hypothesis 7:

[0049] 1.ΔTi v x =x i -x i-1 ΔT i v y =y i -y i-1 ;

[0050] 2. ΔT i v x =x i -x i-1 ;

[0051] 3. ΔT i v y =y i -y i-1 ;

[0052] 4. ΔT i v x =x i -x i-1 ΔT i v y =y i -y i-1 ;

[0053] 5. Assume that the vehicle 102 and the tracked object have the same motion (e.g., curvature). ΔT i v x =x i -x i-1 ΔT i v y =y i -y i-1 ;

[0054] 6. Assume a radial motion of vehicle 102 and the tracked object (e.g., the object's velocity vector points towards the center of the front bumper of vehicle 102). ΔT i v x =x i -x i-1 ΔT i v y =y i -y i-1 .

[0055] 7. Assume there is a cross-radial motion between vehicle 102 and the tracked object (e.g., the object's velocity vector points perpendicular to the radial direction). ΔT i vx =x i -x i-1 ΔT i v y =y i -y i-1 .

[0056] The assumption in assumption 5 listed above is that the objects move on the same road as vehicle 102, and therefore they move with the same motion or curvature. Figure 3-3 Scene 330-2 is shown, in which vehicle 102 has circular motion and object 336 is detected in the field of view. The motion direction or curvature of the main vehicle 102 can be obtained from a vehicle state estimator (VSE) (e.g., executed by the controller of vehicle 102). Based on the VSE of vehicle 102, the motion direction α of object 336 can be easily calculated based on geometry (e.g., α = γ = arcsin[x]). i *curvature]).

[0057] At position 304, the initial velocity for each hypothesis in the hypothesis set is determined. For example, initializer 120 determines the initial velocity for each of the multiple hypotheses 212. As shown below, based on equations 3 and 4, the LLS problem is assigned as a weighted LLS to each hypothesis, where diag[·] refers to the corresponding diagonal matrix:

[0058]

[0059]

[0060] The term (v) in equation 3 x1 ,v y1 Assuming the track has the same motion (curvature) as vehicle 102, the velocity calculated based on the assumption that the track moves radially is (v... x2 ,v y2 The velocity calculated based on the assumption of tracking movement along the intersecting radial direction is (v). x3 ,v y3 ). Item (v x1 ,v y1 ), (v x2 ,v y2 ) and (v x3 ,v y3 () is used only for regularization. In practice, these items can have smaller weights when compared to other information. The actual values ​​of these weights w 3n-1 w 3n w 3n+1 w 3n+2 w 3n+3 w3n+4 It is approximately one-tenth the minimum weight of any other information.

[0061] Equation 4 shows data from two consecutive periods. The weights are the reciprocals of the variance and can be obtained from sensor specifications or data mining. Equation 4 shows that solving for different hypotheses is equivalent to setting different weights to zero while keeping the rest of the weights. For example, Hypothesis 2 listed above is equivalent to setting the terms w5, w6, w7, w8, w9, and w in Equation 4 to zero. 10 The values ​​are reset to zero, and terms w1 and w2 are set to the reciprocals of the variance of the distance change rate. Terms w3 and w4 are also calculated based on the variances of distance and azimuth. For in This is the variance of the distance change rate from the sensor specifications. For w3 and all weights related to longitudinal position change, it can be calculated as the reciprocal of the sum of the longitudinal position variances of two consecutive scans. in It is the azimuth variance from the sensor specifications. It is the rate variance derived from the sensor specifications. This is an adjustable constant representing the uncertainty of the scattering center of the radar point cloud. Similarly, w4 and all weights related to lateral position variation can be calculated by the reciprocal of the sum of the lateral position variances of two consecutive scans.

[0062] The velocity for each hypothesis can be obtained by solving Equation 3. The position can also be estimated using the velocity solved from Equation 3, as well as Equations 5 and 6, which assume that the velocity is nearly constant during the scan period (e.g., four or more consecutive cycles for each hypothesis).

[0063]

[0064]

[0065] At 306, the tracking error associated with each of the hypotheses is determined. For example, initializer 120 determines a first associated level of evidence for each of the multiple hypotheses 212, supporting the initial velocity of that hypothesis. The level of evidence is based on a tracking error score, which is calculated using the cumulative rate of change of position and distance from one or more points of point cloud sensor data 114 associated with one of the tracks in tracking 116. The tracking error score can be determined according to Equation 7. In Equation 7, Δx r It is the positional error (e.g., azimuth) in the distance direction, Δx o It is the positional error in the direction perpendicular to the distance direction. This is the assumed distance change rate error between the data and one or more relevant points from the point cloud sensor data 114. If multiple points from the point cloud sensor data 114 are related to the tracking, the average position and distance change rate can be used. β r β o β rr It is the weight of the difference error.

[0066]

[0067] At this step, the tracking error score is calculated using the latest radar point cloud.

[0068] At 308, another hypothesis is associated with each hypothesis by fusing all initially created hypotheses into a single fused hypothesis. For example, initializer 120 generates the fused hypothesis by combining multiple hypotheses 212 into a single hypothesis. The initial velocity and first associated level of evidence for the fused hypothesis can be based on a set of initial velocities and first associated levels of evidence for the multiple hypotheses. Initializer 120 then includes the fused hypothesis among the multiple hypotheses 212.

[0069] For example, all seven hypotheses from 1 to 7 can be combined to create a new hypothesis, which is called Hypothesis 8. Hypothesis 8 is determined according to Equations 8 and 9. In Equation 8, the term... The i-th hypothesis is generated at position 302, and in Equation 9, the weight of this hypothesis is calculated as evidence E. The term X in Equation 8... F It is a fused hypothesis obtained by summing the weighted hypotheses (e.g., hypothesis 8).

[0070]

[0071]

[0072] After step 308, it is assumed that the initialization time period has ended, and all eight hypotheses are generated using the initial position, initial velocity, and initial tracking error score.

[0073] Assume a tracking period (e.g., steps 310 to 324).

[0074] Assume the tracking period includes steps 310 to 324, and by definition, assume the tracking period is the time after the assumed initialization period and before the tracking is set to a mature state and finally initialized. Before the new tracking in the final initialized and mature tracking 116 is output, the new tracking is similarly tracked, but not reported in tracking 116, until the new tracking is stable and ready for reporting.

[0075] At 310, the best hypothesis is selected from the hypotheses to estimate the velocity measurement of the track. The provisional tracking state of a new track that achieves tracking during this time period can be set as the best track to date (e.g., the hypothesis with the lowest tracking error score). For example, to achieve tracking until maturity, initializer 120 initializes the velocity measurement of each new track before each new track can enter the hypothetical tracking period of that track. In response to including the fused hypotheses among multiple hypotheses 212, initializer 120 selects a first best hypothesis based on a first associated level of evidence for the multiple hypotheses 212.

[0076] At point 312, the hypothesis is updated over time. During the hypothesis tracking period, each hypothesis can be tracked using a Kalman filter with a constant velocity motion model, allowing initializer 120 to solve for each hypothesis with respect to the current period of the hypothesis tracking time. Initializer 120 updates the initial velocity of each of the multiple hypotheses 212 over time.

[0077] At 314, the tracking score associated with each hypothesis is updated. During the hypothesis tracking period, data association is performed on the tracking to determine the tracking error score for each hypothesis. For example, before accumulating the tracking error score for each hypothesis, the tracking error score for each hypothesis can be calculated using Equation 7 using data association between the tracking and one or more associated points from point cloud sensor data 114. For example, initializer 120 determines a second associated level of evidence for each of the multiple hypotheses 212, based on the updated initial velocity, supporting that hypothesis.

[0078] At point 316, any hypothesis that does not meet the evidence threshold is terminated. For example, initializer 120 eliminates any hypothesis from the plurality of hypotheses 212 that has a value for a second associated level of evidence that does not meet the evidence threshold. The evidence threshold includes an evidence ratio calculated for each of the plurality of hypotheses. The evidence ratio calculated for each of the plurality of hypotheses can be a unique evidence ratio among all the plurality of hypotheses. The evidence ratio for each hypothesis can be determined according to Equation 10:

[0079]

[0080] If the assumed evidence ratio ρ Ei If the evidence threshold is less than the initialization threshold, initializer 120 terminates the hypothesis. The evidence threshold can be determined using equation (11):

[0081]

[0082] In equation 11, N 有效This represents the number of hypotheses that remain valid (i.e., have not been terminated) during the current period of the hypothetical tracking timeframe. If no hypothesis has less than the evidence threshold ρ... th If the evidence ratio is less than 2*ρ, then initializer 120 can attempt to terminate the single hypothesis with the minimum evidence ratio; however, this can be done if its evidence ratio is less than 2*ρ. th As a condition. If terminating the hypothesis with the minimum ratio of evidence would be inappropriate, then no hypothesis is terminated during the current period of the hypothesis tracking time.

[0083] At point 318, it is determined whether there is sufficient evidence from the hypotheses to initialize the tracking. In other words, initializer 120 checks whether there is sufficient evidence to finally initialize the tracking using one of the hypotheses. Initializer 120 can use one or more criteria to check whether the evidence for any hypothesis is sufficient to complete the velocity measurement initialization. One criterion could be the assumption that the tracking time period has ended. The duration of the assumed tracking time period can depend on the specific application; however, since this time period adds delay to the tracking initialization, the appropriate duration may vary depending on the vehicle or implementation. For example, the duration is equivalent to eight radar scans in a 20 Hz system. Another criterion could be: only two hypotheses remain valid; all hypotheses except for the two have been terminated so far. A third criterion could be: only three hypotheses exist, and all three hypotheses have an evidence ratio ρ greater than 0.3. Ei This indicates that all three hypotheses produce similar results. When sufficient evidence is determined at 318 (e.g., at least one criterion is met), process 300 proceeds to 324; when insufficient evidence is determined at 318, process 300 proceeds to 320.

[0084] At 320, the remaining hypotheses are measured and updated. For example, in response to identifying two or more remaining hypotheses among multiple hypotheses 212 after eliminating any hypothesis with a value that does not meet the evidence threshold for a second associated level of evidence, initializer 120 measures and updates the predicted movement of the object for each of the remaining multiple hypotheses 212.

[0085] As indicated above at 312, during the hypothesis tracking period, each hypothesis can be tracked using a Kalman filter with a constant velocity motion model, which enables initializer 120 to solve for each hypothesis for the current cycle time of the hypothesis tracking period. At 320, each hypothesis is measured and updated using information derived from the point cloud sensor data 114 associated with the tracking.

[0086] At 322, similar to step 310, the best hypothesis is selected from the hypotheses to serve as the tracking state for a new tracking, enabling further tracking during the hypothesis tracking period. This temporary tracking state allows data association to occur during the hypothesis tracking period. For example, in response to determining two or more remaining hypotheses 212 after eliminating any hypotheses with values ​​that do not meet the second associated evidence level threshold, initializer 120 selects a second best hypothesis based on the second associated evidence level of the remaining multiple hypotheses 212 to replace the previously selected first best hypothesis.

[0087] At 324, the optimal hypothesis is used to initialize the velocity measurement of the tracking. If the condition at 318 indicates that the evidence is sufficient, the optimal hypothesis will be used to initialize the tracking state. For example, in response to determining that only two hypotheses remain out of multiple hypotheses 212 after eliminating any hypotheses with values ​​that do not meet the evidence threshold of the second associated evidence level, the initializer 120 selects a third optimal hypothesis based on the second associated evidence level of the remaining two hypotheses to replace the previously selected first or second optimal hypothesis.

[0088] During the assumed tracking period, new tracking is measured and updated, and the tracking error is recalculated until the new tracking matures, at which point it is subsequently reported in tracking 116. A flag can prevent new tracking from being included and output in tracking 116. A new tracking is added to tracking 116 when the flag indicates that it has been initialized and is otherwise ready to be included in the output of sensing system 104. At 324, object tracker 118 and / or initializer 120 can change the flag to indicate that new tracking has matured and is ready to be included in tracking 116. For example, object tracker 118 can cause sensing system 104 to output tracking of objects (including velocity parameters initialized to a third best assumption) for vehicle-based system 210.

[0089] Example Results

[0090] Figure 4Plot 400 shows the velocity measurement initialization over a period of time during which a stable radar tracking velocity was initialized using multiple assumptions. The results of plot 400 are derived from performance testing of a sensing system (such as sensing system 104) as the object crosses a straight intersecting path along the main vehicle's travel path. In plot 400, lines 402 and 404 show the longitudinal and lateral velocities reported by the Global Positioning System (GPS), respectively. Lines 402 and 404 represent the true velocity of the object. When initialized using a previous tracking velocity initialization technique (i.e., the old method, without using multiple assumptions), lines 410 and 412 track the longitudinal and lateral velocities, respectively. In contrast, lines 406 and 408 demonstrate the significant advantages of using multiple assumptions to initialize velocity measurements (i.e., the new method). Lines 406 and 408 show the longitudinal and lateral velocities tracked using multiple assumptions, which allows the tracked velocity to stabilize to values ​​closer to the true velocity in a shorter time and / or with a smaller variance (e.g., lines 402 and 404), resulting in more stable results than lines 410 and 412. When the object travels slightly perpendicular to the path of the main vehicle, comparing lines 406 and 408 with lines 410 and 412 shows that the example perception system can achieve speed tracking initialization as fast or almost as fast as other speed initialization techniques, but with much higher accuracy than lines 402 and 404.

[0091] Further examples

[0092] Further examples of the above technologies include:

[0093] Example 1. A system comprising a processor configured to: acquire point cloud sensor data indicating signal return reflected from an object in an environment; and establish tracking of an object in the environment using the point cloud sensor data, including initializing velocity measurements of the object by: creating multiple hypotheses for predicted movement of the object; determining, for each of the multiple hypotheses, an initial velocity and a first associated level of evidence supporting that initial velocity; generating a fused hypothesis by combining the multiple hypotheses, the initial velocity of the fused hypothesis and the first associated level of evidence being based on a set of the initial velocities and the first associated levels of evidence for the multiple hypotheses; including the fused hypothesis among the multiple hypotheses; and, in response to including the fused hypothesis among the multiple hypotheses, selecting a first best hypothesis based on the first associated levels of evidence for the multiple hypotheses; updating the initial velocity of each of the multiple hypotheses over time; and determining, for each of the multiple hypotheses, an updated initial velocity supporting that hypothesis. The second associated level of evidence for speed; eliminating any hypothesis from a plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold; determining, in response to eliminating any hypothesis that has a second associated level of evidence that does not meet the evidence threshold, that more than two hypotheses remain from the plurality of hypotheses: updating the predicted movement of the object for each of the remaining hypotheses; and selecting a second best hypothesis to replace the previously selected first best hypothesis based on the second associated level of evidence for the remaining hypotheses; and determining, in response to eliminating any hypothesis that has a second associated level of evidence that does not meet the evidence threshold, that only two hypotheses remain from the plurality of hypotheses: selecting a third best hypothesis to replace the previously selected first best hypothesis or second best hypothesis based on the second associated level of evidence for the remaining two hypotheses; and outputting tracking of the object for the vehicle system, the tracking including a speed parameter initialized to the third best hypothesis.

[0094] Example 2. A system of any other example, wherein the processor is further configured to, before only two hypotheses remain: update the initial velocity of each of the multiple hypotheses over time; for each of the multiple hypotheses, determine a second associated level of evidence supporting the updated initial velocity; and eliminate from the multiple hypotheses any hypothesis that has a value of the second associated level of evidence that does not meet the evidence threshold.

[0095] Example 3. Any other example system, wherein, in further response to determining that the hypothesis tracking period has ended, a second best hypothesis is selected to replace the previously selected first best hypothesis.

[0096] Example 4. Any other example system, where it is assumed that the tracking time period includes multiple frames of a radar system from which point cloud sensor data is obtained.

[0097] Example 5. Any other example system, where it is assumed that the tracking time period includes approximately fifteen frames of the radar system.

[0098] Example 6. Any other example system in which the processor is configured to determine, for each of a plurality of hypotheses, a first associated level of evidence supporting an initial velocity by determining the cumulative rate of change error of position and distance of one or more points in the point cloud sensor data.

[0099] Example 7. Any other example system in which the processor is configured to use a constant motion model to make time updates or measurement updates to a variety of hypotheses.

[0100] Example 8. Any other example system where the evidence threshold comprises an evidence ratio calculated for each of a plurality of hypotheses.

[0101] Example 9. Any other example of a system where the evidence ratio calculated for each of the multiple hypotheses includes the unique evidence ratio among all the multiple hypotheses.

[0102] Example 10. Any other example system where the point cloud sensor data includes point cloud radar data.

[0103] Example 11. A method comprising: acquiring point cloud sensor data from a radar system by an object tracker, the point cloud sensor data indicating radar echoes reflected from an object in an environment; and establishing tracking of an object in the environment using the point cloud sensor data, including initializing a velocity measurement of the object by: creating multiple hypotheses for predicted movement of the object; determining an initial velocity and a first associated level of evidence supporting the initial velocity for each of the multiple hypotheses; generating a fused hypothesis by combining the multiple hypotheses, the initial velocity of the fused hypothesis and the first associated level of evidence being based on a set of the initial velocities of the multiple hypotheses and the first associated level of evidence; including the fused hypothesis among the multiple hypotheses; and, in response to including the fused hypothesis among the multiple hypotheses, selecting a first optimal hypothesis based on the first associated level of evidence of the multiple hypotheses for initializing the velocity measurement of the object.

[0104] Example 12. A method of any other example, further comprising: updating the initial velocity of each of a plurality of hypotheses over time; determining, for each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity; eliminating from the plurality of hypotheses any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold; in response to eliminating any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold, determining that two or more hypotheses remain from the plurality of hypotheses: updating the predicted movement of the object for each of the remaining hypotheses; and selecting a second best hypothesis to replace the previously selected first best hypothesis based on the second associated level of evidence of the remaining hypotheses.

[0105] Example 13. A method of any other example, further comprising: updating the initial velocity of each of a plurality of hypotheses over time; determining, for each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity; eliminating from the plurality of hypotheses any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold; in response to eliminating any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold, determining that only two hypotheses remain from the plurality of hypotheses: selecting a third best hypothesis to replace the previously selected first best hypothesis or second best hypothesis based on the second associated levels of evidence of the remaining two hypotheses; and outputting for the vehicle system a tracking of an object, the tracking including a velocity parameter initialized to the initial velocity of the third best hypothesis.

[0106] Example 14. A method of any other example, further comprising: updating the initial velocity of each of a plurality of hypotheses over time; determining, for each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity; eliminating from the plurality of hypotheses any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold; in response to eliminating any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold, determining that two or more hypotheses remain from the plurality of hypotheses: for each of the remaining hypotheses, updating the predicted movement of the object by measurement; and selecting a second best hypothesis to replace the previously selected first best hypothesis based on the second associated level of evidence of the remaining hypotheses; and in response to eliminating any hypothesis having a value of the second associated level of evidence that does not meet the evidence threshold, determining that only two hypotheses remain from the plurality of hypotheses: selecting a third best hypothesis to replace the previously selected first best hypothesis or second best hypothesis based on the second associated level of evidence of the remaining two hypotheses; and outputting for the vehicle system a tracking of the object, the tracking including a velocity parameter initialized to the third best hypothesis.

[0107] Example 15. A method for any other example, wherein, in further response to determining that the hypothesis tracking period has expired, a second best hypothesis is selected to replace the previously selected first best hypothesis.

[0108] Example 16. Any other example method, where it is assumed that the tracking time period includes multiple frames of the radar system.

[0109] Example 17. Any other example of the method, where the evidence threshold includes a unique evidence ratio calculated for each of the multiple hypotheses.

[0110] Example 18. A method of any other example, wherein the object tracker is configured to determine, for each of a plurality of hypotheses, a first associated level of evidence supporting an initial velocity by determining the cumulative rate of change error of the position and distance of the associated portion of the point cloud sensor data.

[0111] Example 19. Any other example method where the object tracker is configured to use a constant motion model to update time or measurement for multiple hypotheses.

[0112] Example 20. A computer-readable storage medium comprising instructions, which, when executed, cause a processor to perform an object tracker configured to: acquire point cloud sensor data from a radar system, the point cloud sensor data indicating radar echoes reflected from objects in an environment; and establish tracking of objects in the environment using the point cloud sensor data, including initializing velocity measurements of the objects by: creating at least two hypotheses for predicted movement of the objects; determining an initial velocity and a first associated level of evidence supporting the initial velocity for each of the at least two hypotheses; generating a fused hypothesis by combining the at least two hypotheses, the initial velocity of the fused hypothesis and the first associated level of evidence being based on a set of the initial velocity of the at least two hypotheses and the first associated level of evidence; including the fused hypothesis among the at least two hypotheses; and, in response to including the fused hypothesis among the at least two hypotheses, selecting a single optimal hypothesis based on the first associated level of evidence for the at least two hypotheses for initializing velocity measurements of the objects.

[0113] Example 21. A system comprising means for performing the method of any of the preceding examples.

[0114] Example 22. A system including a processor configured to perform a method of any of the preceding examples.

[0115] Example 23. A computer-readable medium comprising instructions that, when executed, cause a processor to perform any of the methods described in the preceding examples.

[0116] Conclusion

[0117] While various embodiments of the present disclosure have been described in the foregoing description and illustrated in the accompanying drawings, it should be understood that the present disclosure is not limited thereto, but can be practiced in various ways within the scope of the following claims. It will be apparent from the foregoing description that various modifications can be made without departing from the scope of the present disclosure as defined by the following claims. In addition to radar systems, problems associated with the initialization of stable radar tracking speeds can arise in other systems (e.g., imaging systems, lidar systems, ultrasonic systems) that identify and process tracking from various sensors. Therefore, although described as improving radar tracking, the techniques described above can be adapted and applied to other problems to efficiently detect and track objects in a scene using other types of object trackers.

[0118] Unless the context explicitly states otherwise, the use of "or" and grammatically related terms indicates an unrestricted, non-exclusive alternative. As used herein, the phrase referring to "at least one" of a list of items means any combination of those items, including a single member. As an example, "at least one of a, b, or c" is intended to cover: a, b, c, ab, ac, bc, and abc, as well as any combination with multiple identical elements (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbc, cc, and ccc, or any other ordering of a, b, and c).

Claims

1. A system comprising a processor configured to: Obtain point cloud sensor data, which indicates signals reflected back from objects in the environment; and Using the point cloud sensor data to establish tracking of objects in the environment includes initializing velocity measurements of the objects in the following manner: Create multiple hypotheses for predicting the movement of the object; A tracking error score is determined for each of the plurality of hypotheses, the tracking error score being calculated as the cumulative rate of change error of position and distance associated with one or more points of point cloud sensor data associated with the tracking; For each of the multiple hypotheses, an initial velocity and a first associated level of evidence supporting that initial velocity are determined, the first associated level of evidence corresponding to the tracking error score of that hypothesis; A fused hypothesis is generated by combining the multiple hypotheses, and the initial velocity and first associated level of evidence of the fused hypothesis are based on the set of the initial velocity and first associated level of evidence of the multiple hypotheses. The fused hypothesis is included among the plurality of hypotheses; as well as In response to including the fused hypothesis among the plurality of hypotheses, a first optimal hypothesis is selected based on a first associated level of evidence for the plurality of hypotheses, the first optimal hypothesis having the lowest associated tracking error score among the plurality of hypotheses; The initial velocity of each of the various assumptions is updated over time; For each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity is determined, the second associated level of evidence corresponding to the updated tracking error score of the hypothesis; Eliminate any hypothesis among the plurality of hypotheses that has a value of a second associated level of evidence that does not meet the evidence threshold; In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, determining that two or more hypotheses remain from the plurality of hypotheses: For each of the remaining multiple assumptions, the predicted movement of the object is measured and updated; and Based on the second associated level of evidence among the remaining multiple hypotheses, a second best hypothesis is selected to replace the previously selected first best hypothesis, the second best hypothesis having the lowest associated tracking error score among the remaining multiple hypotheses after the measurement update; as well as In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, it is determined that only two hypotheses remain among the plurality of hypotheses: Based on the second associated level of evidence for the remaining two hypotheses, a third best hypothesis is selected to replace the previously selected first best hypothesis or second best hypothesis, the third best hypothesis having the lowest associated tracking error score among the remaining two hypotheses; as well as The tracking of the object is output to the vehicle system, the tracking including a velocity parameter initialized to an initial velocity of the third best assumption.

2. The system as described in claim 1, characterized in that, The processor is further configured to, before only two assumptions remain: The initial velocity of each of the various assumptions is updated over time; For each of the multiple hypotheses, determine a second associated level of evidence supporting the updated initial velocity; as well as Eliminate any hypothesis from the plurality of hypotheses that has a value of a second associated level of evidence that does not meet the evidence threshold.

3. The system as described in claim 1, characterized in that, In response to determining that the hypothesis tracking period has ended, a second best hypothesis is selected to replace the previously selected first best hypothesis.

4. The system as described in claim 3, characterized in that, The assumed tracking time period includes multiple frames of the radar system from which the point cloud sensor data is obtained.

5. The system as described in claim 4, characterized in that, The assumed tracking time period includes approximately fifteen frames from the radar system.

6. The system as described in claim 1, characterized in that, The processor is configured to determine, for each of the plurality of hypotheses, a first associated level of evidence supporting the initial velocity of that hypothesis by determining the cumulative rate of change error of position and distance at one or more points of the point cloud sensor data.

7. The system as described in claim 1, characterized in that, The processor is configured to use a constant motion model to update the various hypotheses over time or by measurement.

8. The system as described in claim 1, characterized in that, The evidence threshold includes the evidence ratio calculated for each of the multiple hypotheses.

9. The system as described in claim 8, characterized in that, The evidence ratio calculated for each of the multiple hypotheses includes the unique evidence ratio among all the multiple hypotheses.

10. The system as claimed in claim 1, characterized in that, The point cloud sensor data includes point cloud radar data.

11. A method, the method comprising: Point cloud sensor data is obtained from a radar system by an object tracker, and the point cloud sensor data indicates radar echoes reflected from objects in the environment; as well as Using the point cloud sensor data to establish tracking of objects in the environment includes initializing velocity measurements of the objects in the following manner: Create multiple hypotheses for predicting the movement of the object; A tracking error score is determined for each of the plurality of hypotheses, the tracking error score being calculated as the cumulative rate of change error of position and distance associated with one or more points of point cloud sensor data associated with the tracking; For each of the multiple hypotheses, an initial velocity and a first associated level of evidence supporting that initial velocity are determined, the first associated level of evidence corresponding to the tracking error score of that hypothesis; A fused hypothesis is generated by combining the multiple hypotheses, and the initial velocity and first associated level of evidence of the fused hypothesis are based on the set of the initial velocity and first associated level of evidence of the multiple hypotheses. The fused hypothesis is included among the plurality of hypotheses; as well as In response to including the fused hypothesis among the plurality of hypotheses, a first optimal hypothesis is selected based on a first associated level of evidence of the plurality of hypotheses to initialize the velocity measurement of the object, the first optimal hypothesis having the lowest associated tracking error score among the plurality of hypotheses.

12. The method of claim 11, further comprising: The initial velocity of each of the various assumptions is updated over time; For each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity is determined, the second associated level of evidence corresponding to the updated tracking error score of the hypothesis; Eliminate any hypothesis among the plurality of hypotheses that has a value of a second associated level of evidence that does not meet the evidence threshold; In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, determining that two or more hypotheses remain from the plurality of hypotheses: For each of the remaining multiple assumptions, the predicted movement of the object is measured and updated; and Based on the second associated level of evidence for the remaining multiple hypotheses, a second best hypothesis is selected to replace the previously selected first best hypothesis.

13. The method of claim 11, further comprising: The initial velocity of each of the various assumptions is updated over time; For each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity is determined, the second associated level of evidence corresponding to the updated tracking error score of the hypothesis; Eliminate any hypothesis among the plurality of hypotheses that has a value of a second associated level of evidence that does not meet the evidence threshold; In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, it is determined that only two hypotheses remain among the plurality of hypotheses: Based on the second associated level of evidence for the remaining two hypotheses, a third best hypothesis is selected to replace the previously selected first best hypothesis or second best hypothesis, the third best hypothesis having the lowest associated tracking error score among the remaining two hypotheses; as well as The tracking of the object is output to the vehicle system, the tracking including a velocity parameter initialized to an initial velocity of the third best assumption.

14. The method of claim 11, further comprising: The initial velocity of each of the various assumptions is updated over time; For each of the plurality of hypotheses, a second associated level of evidence supporting the updated initial velocity is determined, the second associated level of evidence corresponding to the updated tracking error score of the hypothesis; Eliminate any hypothesis among the plurality of hypotheses that has a value of a second associated level of evidence that does not meet the evidence threshold; In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, determining that two or more hypotheses remain from the plurality of hypotheses: For each of the remaining multiple assumptions, the predicted movement of the object is measured and updated; and Based on the second associated level of evidence among the remaining multiple hypotheses, a second best hypothesis is selected to replace the previously selected first best hypothesis, the second best hypothesis having the lowest associated tracking error score among the remaining multiple hypotheses after the measurement update; as well as In response to eliminating any hypothesis among the plurality of hypotheses that has a second associated level of evidence that does not meet the evidence threshold, it is determined that only two hypotheses remain among the plurality of hypotheses: Based on the second associated level of evidence for the remaining two hypotheses, a third best hypothesis is selected to replace the previously selected first best hypothesis or second best hypothesis, the third best hypothesis having the lowest associated tracking error score among the remaining two hypotheses; as well as The tracking of the object is output to the vehicle system, the tracking including a velocity parameter initialized to an initial velocity of the third best assumption.

15. The method as described in claim 14, characterized in that, In response to determining that the hypothesis tracking period has expired, a second best hypothesis is selected to replace the previously selected first best hypothesis.

16. The method as described in claim 15, characterized in that, The assumed tracking time period includes multiple frames of the radar system.

17. The method as described in claim 14, characterized in that, The evidence threshold includes a unique evidence ratio calculated for each of the multiple hypotheses.

18. The method as described in claim 11, characterized in that, The object tracker is configured to determine a first associated level of evidence supporting an initial velocity for each of the plurality of hypotheses by determining the cumulative rate of change error of position and distance in the associated portion of the point cloud sensor data.

19. The method as described in claim 11, characterized in that, The object tracker is configured to use a constant motion model to update the various hypotheses over time or by measurement.

20. A computer-readable storage medium comprising instructions that, when executed, cause a processor of a radar system to execute an object tracker configured to perform the method of any one of claims 11-19.

21. A radar system comprising the system according to any one of claims 1-10, wherein, The radar system is configured to perform the method of any one of claims 11-19 using the processor of the system.

Images