Target tracking methods using limited PMHT radar measurements
The method addresses inefficiencies in PMHT radar tracking by applying sensor-specific angle restrictions and visibility checks to limit measurements to visible sources, improving tracking accuracy and reducing computational burden.
Patent Information
- Application Number
- DE102023005261
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2023-12-21
- Publication Date
- 2024-12-05
- Estimated Expiration
- 2043-12-21
AI Technical Summary
Existing PMHT radar tracking methods face inefficiencies due to the assumption of independent measurement assignments, leading to potential filter entrapment in local maxima and poor representation of target length using Gaussian distributions, necessitating a more accurate modeling of measurement origins and computational burden reduction.
A method for target tracking with limited PMHT radar measurements is proposed, incorporating sensor-specific angle restrictions and visibility checks to limit measurements to visible sources, using a contour-based model and probabilistic assignment probabilities to enhance tracking accuracy.
The solution provides a robust PMHT radar tracker that reduces computational burden and improves tracking accuracy by restricting measurements to visible sources, avoiding local maxima and enhancing the reliability of target tracking.
Smart Images

Figure 00000000_0000_ABST
Abstract
Description
[0001] The invention relates to a method for target tracking with limited PMHT radar measurements according to the preamble of claim 1.
[0002] J.-F. Bariant, L.L. Palacios, J. Granitzka and H. Groener, "Extended Target Tracking with Constrained PMHT," 2022 25th International Conference on Information Fusion (FUSION), Linköping, Sweden, 2022, pp. 1-8, doi: 10.23919 / FUSION49751.2022.9841232 describes certain problems of extended target tracking. A method is proposed to accurately model the origin of measurements on the target surface. This is achieved by abandoning the usual hypothesis that the independence of the assignment of measurements to possible measurement sources allows the assumption that a certain number of measurements originate from a few specific sources. A Gaussian distribution is a poor representation of the length of a target.A method for discretizing the length is therefore proposed to estimate its distribution without the Gaussian assumption, while avoiding the computational overhead of multi-hypothesis tracking for each target. The efficiency of the implementation is demonstrated on simulated and real radar data.
[0003] The invention is based on the object of providing a novel method for target tracking with limited PMHT radar measurements.
[0004] The object is achieved according to the invention by a method for target tracking with limited PMHT radar measurements having the features of claim 1.
[0005] Advantageous embodiments of the invention are the subject of the subclaims.
[0006] A method for target tracking with restricted PMHT radar measurements is proposed, wherein measurement models of objects determined by measurements from at least one radar sensor are defined by means of a contour having corners and edges as measurement sources. According to the invention, an angle restriction is applied for each sensor, which takes into account the sensor-specific visibility restrictions while considering a rotation angle of the object in a coordinate system of the radar sensor. The rotation angle is determined as the difference between a yaw angle of the object and a sensor-to-object angle. If the rotation angle is greater than a specified upper limit angle, it is concluded that a trailing edge and a leading edge are not visible. Otherwise, if the rotation angle is smaller than a specified lower limit angle, it is concluded that a left edge and a right edge are not visible.Otherwise, it is concluded that no angle constraint exists and the left edge and the right edge are visible. Furthermore, a number of constrained measurements is determined as the product of a number of measurements and a proportion. If the number of constrained measurements is less than a number of visible edges, then no constraints exist. Otherwise, it is determined whether the number of visible edges is 1, in which case one measurement is constrained to each visible edge and one measurement to each corner adjacent to the visible edge, with the remainder of the constraint measurements being assigned to the visible edge, if it is the right edge or the left edge.If the number of visible edges is not equal to 1, it is determined whether the number of visible edges is equal to 2, in which case one measurement is constrained to the common corner of the two visible edges and the rest of the constraint measurements are constrained to the two visible edges, where if the number of visible edges is not equal to 2, no constraints apply.
[0007] According to the invention, a method for object visibility testing and a constraint algorithm are proposed that uses the visibility information to restrict a portion of the assigned measurements to visible measurement sources of the tracked object by using special techniques for generating various challenges.
[0008] The inventive solution provides a robust, limited PMHT radar tracker.
[0009] In advanced object tracking using PMHT (probabilistic multi hypothesis tracker), a sensor-specific constraint algorithm is required for the assignment of measurements, since the PMHT hypothesis, which states that all assignments are independent of each other, can cause the filter to get stuck in unwanted local maxima.
[0010] Therefore, an algorithm is proposed to restrict the measurements of the constraints to the objects tracked by radar, taking into account the object's visibility and the number of measurements associated with the object. In principle, some measurements associated with visible sides should be restricted, while the well-known PMHT algorithm allows assignment to all existing measurement sources with equal probability.
[0011] Embodiments of the invention are explained in more detail below with reference to drawings.
[0012] Showing: Fig. 1 a schematic view of a proposed measurement model of an object, Fig. 2 a schematic view illustrating the object visibility of several objects for a sensor, Fig. 3 a schematic view illustrating the angular limitation of the object visibility of the object for the sensor, Fig. 4 a schematic view of an algorithm for determining an angle constraint, and Fig. 5 a schematic view of a constraint algorithm.
[0013] Corresponding parts are provided with the same reference numerals in all figures.
[0014] In the context of the present invention, the following terminology is used: Measurement source: A measurement source represents a location on a body of a target vehicle from which a measurement originates by means of a sensor signal.
[0015] Measurement: A measurement is a sensor signal that originates from an object in the environment observed by a sensor (example: distance from the sensor to a point on the object) and is assigned to a specific measurement source.
[0016] Restricted measurement: A restricted measurement is a measurement with an assignment and / or restriction to a specific measurement source.
[0017] State vector: A state vector describes an object according to a given model. In PMHT, the state includes a position, a yaw angle (the object's orientation), a yaw rate (the rate of change in the object's orientation), latitude, and longitude. PMHT: Probabilistic Multiple Hypothesis Tracker
[0018] Clutter: A measurement that is not related to the tracked object, such as ground detections.
[0019] Inside: A measurement source that represents a body area of the object, omitting a contour area (example: roof, chassis).
[0020] Fig. 1 is a schematic view of a proposed measurement model 1 of an object 6, in particular a vehicle. A contour 2 or bounding box of the measurement model 1 is defined by corners FL, FR, RR, RL and edges 4.1, 4.2, 4.3, 4.4, which extend outside the object 6, in particular also outside its side mirrors. Furthermore, an interior region 5 of the measurement model 1 is defined within the contour 2, which is drawn more closely around the object 6, so that protruding areas of the object 6, for example the side mirrors, can be located outside the interior region 5. Furthermore, clutter can occur. The corners FL, FR, RR, RL, edges 4.1, 4.2, 4.3, 4.4, the interior region 5, and the clutter can represent measurement sources.
[0021] PMHT uses ten vehicle measurement sources, including the four corners FL, FR, RR, RL, the four edges 4.1, 4.2, 4.3, 4.4, the interior area 5, and the clutter data.
[0022] For measurement clusters, an assignment probability is calculated for each measurement source. This probability expresses how likely it is that a measurement originates from a specific object. This information is used for a weighted update.
[0023] The object is represented as a state vector that is tracked by a suitable Kalman filter.
[0024] According to the present invention, it is proposed to use the visibility information to increase the probability for some measurements to have been caused by visible measurement sources compared to other hidden or concealed measurement sources.
[0025] For example, a radar tracker can be provided that uses a radar point cloud as an input signal to track extended objects such as motor vehicles or bicycles. In principle, a PMHT tracker models the object surfaces into a number of possible measurement sources, for example, the right side, the left corner, etc. The PMHT assigns the clusters of input signals from the radar detections to the correct measurement source in a probabilistic manner.
[0026] Like any other optimization algorithm, PMHT can get stuck in local maxima. Therefore, some constraints are applied to the assignment to different measurement sources on the object surface. These constraints are essentially based on the object's visibility.
[0027] For example, if an object in front is measured from its back side, then more measurements can be expected from that back side. A constraint can be applied as follows: At least n measurements should originate from the visible back side with a certain probability Pc, which we call the probability that the constraint is correct.
[0028] The present invention provides a robust visibility algorithm and selects the correct mapping constraints based on this visibility information.
[0029] The solution is described below in overview: Step 1: - Determine which sides or edges 4.1, 4.2, 4.3, 4.4 or corners FL, FR, RR, RL are visible (for example, a front edge 4.1 points to its own sensor), - Enhance the visibility information by adding an angle check to ensure that the edge 4.1, 4.2, 4.3, 4.4 itself is easily observable and has a high angle of incidence with the sensor beams. Step 2: - Using the visibility information (which edge 4.1, 4.2, 4.3, 4.4 or corner FL, FR, RR, RL is clearly observable (visible) in the field of view (FOV) of the sensor) to choose the correct mapping constraints, - The choice of constraints depends on the number of visible edges 4.1, 4.2, 4.3, 4.4 (two edges and one vertex or one edge and two vertices), - The number of measurements to be restricted for each visibility model is an adjustment ratio of the total number of cluster measurements. Step 3: - Use of the constrained PMHT mechanism to incorporate the chosen assignment constraints into (posterior) weighting factors (weighting factors that reflect the probability that a particular measurement originates from a particular measurement source).
[0030] Fig. Figure 2 is a schematic view illustrating the object visibility of several objects 6.1 to 6.5, for example, vehicles located in a coordinate system spanned by the spatial directions x and y. For each object 6.1 to 6.5, a front edge 4.1, a right edge 4.2, a rear edge 4.3, a left edge 4.4, a front left corner FL, a front right corner FR, a rear right corner RR, and a rear left corner RL are defined.
[0031] A sensor 7, for example a radar sensor 7, at the coordinate origin of the coordinate system perceives the objects 6.1 to 6.5 and at least some of their edges 4.1 to 4.4 and corners FL, FR, RR, RL from different viewing angles 8 due to their different positions and orientations. The spatial direction x can denote a longitudinal direction of a vehicle on which the sensor 7 can be arranged. The spatial direction y can denote a transverse direction of the vehicle.
[0032] In principle, with PMHT, an assignment probability can be calculated for each existing object and measurement source for each measurement and measurement source used as an input signal. Clusters of measurements can initially be assigned to specific objects.
[0033] For each measurement in a cluster, an assignment probability for each measurement source should be calculated. An important input for calculating the probability is the visibility of the measurement sources. It is generally assumed that measurement sources invisible to the sensor have lower assignment probabilities.
[0034] Corners are only considered visible if the two enclosing sides (edges) of the object can be perceived by the sensor. Accordingly, if only a single side (edge) is visible, it cannot be assumed that the two corners are also visible.
[0035] The visibility of the measurement sources must be calculated from the sensor's perspective. This can be done by applying the cross product.
[0036] The sign (positive or negative) of a cross product (x) between each pair of adjacent vertices FL, FR, RR, RL indicates whether the edge 4.1 to 4.4 located between the pair of vertices FL, FR, RR, RL is visible to the sensor 7 or not.
[0037] For example, for object 6.1: leading edge 4.1: FL × FR = -ve right edge 4.2: FR × RR = -ve trailing edge 4.3: RR × RL = +ve left edge 4.4: RL × FL = +ve
[0038] Where -ve means not visible and +ve means visible.
[0039] The rear edge 4.3 and the left edge 4.4 are therefore visible to sensor 7.
[0040] Fig. Figure 3 is a schematic view illustrating the angular constraint on the visibility of object 6.3 for sensor 7. The orientation of object 6.3 is determined by a yaw angle α (specifically, an angle around a vertical axis) and a sensor-to-object angle β, from which a rotation angle γ can be calculated. The yaw angle α is estimated by tracking over time, supported by sensor measurements.
[0041] The visibility of the object's measurement sources can be further restricted by performing the visibility check with an angle-based algorithm. This leads to more conservative results. In particular, it takes into account the fact that even a slightly visible side (edge) of an object is classified as visible by a cross-product-based check. Nevertheless, it can be assumed that a barely visible side will not provide any measurements or will not provide reliable measurements.
[0042] For the test, the sensor-specific visibility limits, the object orientation (yaw angle α) in a fixed coordinate system, and a sensor-to-object angle β (the angle of the line of sight from the sensor to the object) are taken into account. The rotation angle γ of the object is the smallest absolute angle between the line of sight and the object's principal axis, which is given by the yaw angle α. The rotation angle γ is therefore defined for a range from 0° to 90°. The rotation angle γ can be calculated in such a way that the classification is symmetrical for both the back and front, as well as for the left and right sides. To determine lateral visibility, angular thresholds (limit angle ϑ) are used, which are also defined in the range from 0° to 90° defined for the rotation angle γ. Typical values for the limit angle ϑ are 30° for the lower limit angle ϑl. ower and 60° for the upper limit angle ϑupper These values are therefore 1 / 3 and 2 / 3, respectively, of the 90° interval in which the rotation angle γ is defined. Since the rotation angle γ of the object is the absolute value of the smallest angle between the line of sight and the object's principal axis, only these two values need to be used.
[0043] In Example 1, the rotation angle γ is less than 30°. Therefore, the left and right edges are assumed to be hidden. They are therefore unlikely to reflect measurement signals. The cross-product-based visibility check reveals that the left and back edges, including the enclosed corner, are visible. The angle-based visibility check reveals that the left and right edges are classified as hidden, that only the back edge is classified as visible, and that the back left and back right corners are hidden.
[0044] In Example 2, the rotation angle γ is greater than 60°. Therefore, the back and front edges are assumed to be hidden. They are therefore unlikely to reflect measurement signals. The cross-product-based visibility check reveals that the left and right edges, including the enclosed corner, are visible. The angle-based visibility check reveals that the front and back edges are classified as hidden, that only the left edge is classified as visible, and that the back left corner and the front left corner are hidden.
[0045] For each sensor 7, an angle constraint is applied that takes into account the sensor-specific visibility constraints, taking into account the rotation angle γ of the object 6.3 in the coordinate system of the sensor 7. For example, for object 6.3:
[0046] By applying the cross product based visibility check, it becomes clear that the back edge 4.3 and the left edge 4.4 are visible to the sensor 7.
[0047] The application of the angle constraint shows that the upper threshold for alignment is exceeded and the rear edge 4.3 is hidden or not visible to the sensor 7.
[0048] Therefore, only the left edge 4.4 is visible to sensor 7. The back side is barely visible, but the visibility test classifies it as visible. Since it is barely visible from the sensor's perspective, it cannot be considered a reliable measurement source.
[0049] Fig. Figure 4 is a schematic view of an algorithm for determining the angle constraint. In step S1, the rotation angle γ is determined as the difference between the yaw angle α and the sensor-to-object angle β. γ=α−β
[0050] In a step S2 it is determined whether the rotation angle γ is greater than a specified upper limit angle ϑ upper If this is the case, then in step S3 it is concluded that the trailing edge 4.3 and the leading edge 4.1 are not visible (hidden). Therefore, measurements cannot be assigned to these edges 4.3, 4.1.
[0051] If the rotation angle γ is not greater than the upper limit angle ϑ upper then in a step S4 it is determined whether the rotation angle γ is smaller than a specified lower limit angle ϑ lower If this is the case, then in step S5 it is concluded that the left edge 4.4 and the right edge 4.2 are not visible (hidden). Therefore, measurements cannot be assigned to these edges 4.3, 4.1.
[0052] If the rotation angle γ is not smaller than the lower limit angle ϑl ower, then in a step S6 it is concluded that as a result of the visibility check based on the cross product there is no angle restriction and both edges, in particular the left edge 4.4 and the right edge 4.2, are visible.
[0053] Fig. Figure 5 is a schematic view of a constraint algorithm.
[0054] Given a cluster of measurements associated with an object, the PMHT calculates assignment probabilities for each of the measurement sources (edges, corners, etc.) for each of the measurements.
[0055] Taking visibility into account, the probability of assigning a measurement to hidden measurement sources is reduced. According to the present invention, the assignment is further reduced to a single edge of the object if one of the angle constraint criteria is met.
[0056] In Example 1 above, the back edge and the left edge are clearly visible. Therefore, the assignment probability of the measurements to the back edge, right edge, and back left corner should not be calculated in the conventional way. Instead, measurements from these visible sources are assigned a higher weight.
[0057] The restriction of a measurement to an edge 4.1 to 4.4 or to a corner FL, FR, RR, RL is based on two criteria.
[0058] The constraint measurement should show the highest probability for the considered edge 4.1 to 4.4 or vertex FL, FR, RR, RL compared to the other edges 4.1 to 4.4 and vertices FL, FR, RR, RL.
[0059] The constraint measurement should have the highest Mahalanobis distance to the considered edge 4.1 to 4.4 or corner FL, FR, RR, RL that is less than a specific threshold based on the chi-square gate.
[0060] The number of measurements to be restricted is a proportion r of the total number of measurements assigned to object 6.1 to 6.5.
[0061] In a step S7, a number of limited measurements n c determined as the product of a number of measurements n and the proportion r. nc=r⋅n
[0062] In a step S8 it is determined whether the number of restricted measurements n c is smaller than a number of visible edges s.
[0063] If this is the case, the algorithm is terminated and no restrictions apply. If the allowed maximum number of restricted measurements n cis not less than the number of visible edges s, then in a step S9 it is determined whether the number of visible edges s is equal to 1. If this is the case, then in a step S10 one measurement is restricted to the respective visible edge 4.1 to 4.4 and one measurement is restricted to each corner FL, FR, RR, RL adjacent to the visible edge 4.1, if available. The remainder of the constraint measurements are assigned to the visible edge 4.1 to 4.4 if it is the right edge 4.2 or the left edge 4.4.
[0064] If the number of visible edges s is not equal to 1, then in a step S11 it is determined whether the number of visible edges s is equal to 2. If this is the case, then in a step S12 a measurement is restricted to the common corner FL, FR, RR, RL of the two visible edges 4.1 to 4.4. Then, the remaining constraint measurements (if available) are restricted to the two visible edges 4.1 to 4.4. Each of the two edges 4.1 to 4.4 can have the same number of measurements, or the number of measurements is assigned depending on the length of the two visible edges 4.1 to 4.4.
[0065] If the number of visible edges s is not equal to 2, then the algorithm terminates and there are no restrictions.
[0066] Measurements should not be restricted to sources classified as hidden by the sensor. Therefore, measurements can only be restricted to visible sources.
Claims
[1] Method for target tracking with limited PMHT radar measurements, wherein measurement models (1) of objects (6) determined by measurements of at least one radar sensor (7) are defined by means of a contour (2) having corners (FL, FR, RR, RL) and edges (4.1, 4.2, 4.3, 4.4) as measurement sources, characterized bythat for each sensor (7) an angle restriction is applied which takes into account the sensor-specific visibility restrictions taking into account a rotation angle (γ) of the object (6) in a coordinate system of the radar sensor (7), wherein the rotation angle (γ) is determined as the difference between a yaw angle (α) of the object (6) and a sensor-to-object angle (β), wherein if the rotation angle (γ) is greater than a specified upper limit angle, it is concluded that a rear edge (4.3) and a front edge (4.1) are not visible, otherwise if the rotation angle (γ) is smaller than a specified lower limit angle, it is concluded that a left edge (4.4) and a right edge (4.2) are not visible, otherwise it is concluded that no angle restriction exists and the left edge (4.4) and the right edge (4.2) are visible, wherein a number of restricted measurements is determined as the product of a number of measurements and a proportion, wherein if the number of restricted measurements is less than a number of visible edges (4.1 to 4.4), there are no restrictions, otherwise it is determined whether the number of visible edges (4.1 to 4.4) is equal to 1, in which case one measurement is restricted to the respective visible edge (4.1 to 4.4) and one measurement is restricted to each corner (FL, FR, RR, RL) adjacent to the visible edge (4.1 to 4.4), the rest of the restriction measurements being assigned to the visible edge (4.1 to 4.4), if it is the right edge (4.2) or the left edge (4.4), wherein if the number of visible edges (4.1 to 4.4) is not equal to 1, it is determined whether the number of visible edges (4.1 to 4.4) is equal to 2, where in this case a measurement is taken on the common corner (FL, FR, RR, RL) of the two visible edges (4.1 to 4.4) and the rest of the constraint measurements are restricted to the two visible edges (4.1 to 4.4), whereby if the number of visible edges (4.1 to 4.4) is not equal to 2, no constraints exist. [2] Method according to claim 1, characterized by that if the number of visible edges (4.1 to 4.4) is 2, the same number of measurements is assigned to each of the two visible edges (4.1 to 4.4) or the number of measurements is assigned to the two visible edges (4.1 to 4.4) depending on the length of the two visible edges (4.1 to 4.4). [3] Method according to claim 1 or 2, characterized by that the visibility of an edge (4.1 to 4.4) is determined by forming the cross product of the visibility of the vertices (FL, FR, RR, RL) that bound the respective edge (4.1 to 4.4).
