Shielded dynamic target identification and tracking method based on deep learning

Abstract

The invention relates to sensing and tracking of intelligent driving in rain, snow, fog, night and large shielding scenes. In order to solve the problems of target invisibility, track interruption and misconnection caused by low visibility, a shielded dynamic target identification and tracking method based on deep learning is provided; according to the method, under a unified spatial index, sparse point cloud and road topology, passable and traverse areas, static shielding volume coding, ray marking visibility boundary and shielding entrance and exit are carried out; generating a motion voting field with consistent visibility through multi-agent situation reasoning, performing normal directional enhancement according to a passable boundary, and extracting a risk corridor; voxelization is carried out on the multi-frame point cloud in the corridor, a dynamic sparse voxel map is constructed, a voting field is used for gating cross-frame edge connection, occupation changes and infinitesimal displacement are aggregated, and three-dimensional verification candidates, limited state estimation and a time continuous track are obtained; outputting the target and the corridor to which the target belongs, and forming a potential conflict zone according to the intersection of the target and the own vehicle path; the reproduction rate and the advance are improved, and cross-lane misconnection is reduced.

Description

Technical Field

[0001] This application relates to the field of perception and tracking technology in occluded scenarios, and in particular to a deep learning-based method for recognizing and tracking occluded dynamic targets. Background Technology

[0002] Intelligent driving perception typically relies on sparse point clouds from LiDAR, combined with cameras, millimeter-wave radar, and high-definition maps to reconstruct road topology and traversable areas and identify dynamic targets. In rain, snow, fog, at night, and under conditions with large obstructions, increased noise and visual obstruction significantly impact the early detection and continuous tracking of potential pedestrians crossing or merging vehicles.

[0003] Existing methods mainly fall into two categories: the first is the detection-tracking paradigm, which uses point cloud / image BEV segmentation and 3D bounding boxes, combined with Kalman filtering, data association, and finite map priors to complete target tracking; the second is the occupancy mapping paradigm, which uses point cloud rays to construct (dynamically) occupancy grids, infers accessibility and visibility in free space, and adopts conservative strategies at the planning layer. For occlusion, line-of-sight rays, heuristic region dilation, or confidence decay are often used to characterize the impact, but these methods generally do not provide closed volumetric representations of occlusion structures such as buildings and parked vehicles, and rarely integrate them with lane topology, passability, and pedestrian crossing semantics, resulting in a relatively coarse representation of occluded entrances / exits and potential targets on the invisible side.

[0004] The above methods share common shortcomings in low visibility and large occlusion scenarios: they do not explicitly model visibility boundaries and occlusion volumes, making it difficult to accurately locate occlusion entrances and exits; they lack visibility constraints consistent with road topology to limit the reachability of targets, and cross-frame associations are prone to false connections across lanes; and they lack a confirmation mechanism based on occupancy changes during the invisible-visible transition process, resulting in weak evidence, easy trajectory interruption, and unstable early risk warnings.

[0005] Therefore, how to provide a method for identifying and tracking occluded dynamic targets has become a technical problem that urgently needs to be solved by those in the field. Summary of the Invention

[0006] One objective of this invention is to provide a deep learning-based method for identifying and tracking occluded dynamic targets. This method enables the explicit characterization of line-of-sight visibility and occlusion entrances / exits based on sparse point clouds and road topology, passability, and pedestrian crossing semantics under conditions of rain, snow, fog, night, and large occlusion. It also defines and continuously verifies the actual reachable channels for dynamic targets to enter the visible side from the invisible side in the vehicle coordinate system, thereby supporting continuous tracking and early warning of targets during occlusion transitions.

[0007] A deep learning-based method for identifying and tracking occluded dynamic targets according to an embodiment of the present invention includes:

[0008] S1. On roads with rain, snow, fog, night, and large obstructions, using the vehicle's pose as a reference, sparse point clouds are uniformly encoded with road topology, passable areas, pedestrian crossing areas, and static obstruction volumes. Ray testing is performed on the obstruction volumes, the visibility boundaries and obstruction entrances and exits are marked, and the visible / invisible sides are distinguished to generate scene element maps to limit the reachable space of potential dynamic targets.

[0009] S2. Based on the scene element map of S1, the potential dynamic target is explicitly modeled as an agent using a multi-agent contextual reasoning network. Combining road topology, pedestrian crossing and vehicle flow interaction patterns and occlusion geometry, the existence probability, movement direction and occlusion degree are inferred. A movement voting field with consistent visibility is output. At the occlusion entrances and exits and intersections, the normal direction of the passable boundary is enhanced. A risk corridor is generated in the voting enhancement zone. The risk corridor carries the direction tolerance and occlusion consistency weight.

[0010] S3. Based on the risk corridor of S2, multi-frame sparse point cloud voxels are formed in the corridor to form a dynamic sparse voxel map. Cross-frame edges are gated by a motion voting field with consistent visibility. Only voxels that are in the same direction as the risk corridor and do not cross static occlusion volumes are connected. The edge strength is adjusted according to the voting uncertainty. When the corridor crosses multiple lanes, the edges are independently gated with tolerance in each direction to avoid misconnection between corridors.

[0011] S4. On the dynamic sparse voxel map of S3, aggregate cross-frame occupancy changes and small displacements along the risk corridor to form a chain of micro-motion evidence with consistent direction. Suppress segments that cross static occlusion volumes or deviate from the direction of the risk corridor. Allow short invisible segments for occlusion entrances and exits, but pairwise occupancy changes must appear on both sides of the visibility boundary for confirmation. Output three-dimensional verification candidates and evidence strength.

[0012] S5. Based on the three-dimensional verification candidates of S4, combined with the directional tolerance and occlusion consistency weight of S2, the candidate center, velocity and orientation are constrained and estimated to obtain the initial state estimate and uncertainty interval. The constraints include: the center must not fall into the static occlusion volume or the impassable area, the velocity is mainly projected in the risk corridor direction, and the occlusion consistency weight is used to participate in the weighting to suppress unphysical movements.

[0013] S6. Based on the initial state estimate in S5, a time-continuous trajectory is generated within the risk corridor. When it is temporarily invisible or the number of points decreases sharply, it extends along the direction of the risk corridor based on the motion voting field with consistent visibility. The extension terminates when it reaches the visibility boundary and there is a static occlusion volume in front. When a new micro-motion evidence chain appears in subsequent frames and is consistent with the original risk corridor, the reproduction is completed.

[0014] S7. Based on the trajectory of S6 and the three-dimensional verification candidates of S4, output the target position, speed, orientation and the risk corridor to which it belongs to the vehicle decision-making.

[0015] Optionally, the unified encoding of sparse point clouds with road topology, passable areas, pedestrian crossing areas, and static occlusion volumes specifically involves: in a vehicle coordinate system defined by the vehicle's pose, using a unified spatial index to align the above elements to the same reference frame, and recording semantic labels and adjacency relationships on the same spatial unit; the spatial index is a two-dimensional mesh or a three-dimensional voxel; the semantic labels include road topology identifiers, passable attributes, pedestrian crossing attributes, and occlusion attributes; the adjacency relationships include lane connection relationships, contact relationships between passable area boundaries and static occlusion volumes, to ensure that in rain, snow, fog, nighttime, and large occlusion scenarios, each element can be directly used for occlusion determination and visibility calculation.

[0016] Optionally, the multi-agent contextual reasoning network specifically comprises: using the scene element graph obtained in S1 and the reachability space of potential dynamic targets as input, constructing a graph in the vehicle coordinate system consisting of lane segments, pedestrian crossing areas, occluded entrances / exits, and passable boundary nodes; using the connection relationship of the road topology, the conflict relationship between the pedestrian crossing area and the lane, and the visible / invisible side relationship of the line-of-sight visibility boundary as the edges of the graph; explicitly modeling potential dynamic targets as agents on the above graph, that is, generating agent hypotheses at the nodes of occluded entrances / exits and intersections, the agent's state includes center position candidate, orientation candidate, existence probability, and occlusion degree; the network limits the allowed movement direction of the agents based on the road topology and passable areas, and limits the visibility consistency of the agents based on the occlusion geometry, thereby outputting a set of agents consistent with physical reachability for subsequent movement voting;

[0017] The specific traffic flow interaction rules are as follows: Within the scene element graph, the following relationship along the lane direction within the same lane, the yielding relationship between adjacent lanes at merging / exit points, and the crossing priority relationship between lanes and pedestrian crossing areas at conflict points are expressed in a regularized manner. The above relationships are mapped as directed constraints of the graph, which are used to restrict the relative direction and passability order between agents in the multi-agent contextual reasoning network.

[0018] The directional enhancement based on the normal of the accessible boundary specifically involves: within the neighborhood of the obstructed entrance / exit and the intersection, directional gain is applied to the voting components pointing towards the visible side based on the local normal of the accessible boundary; for the obstructed entrance / exit, the normal pointing from the invisible side to the visible side is taken as the priority direction; for the intersection, the bidirectional normal corresponding to the conflict line is taken as the priority direction; the directional enhancement only applies to the voting components located in the reachable space of the potential dynamic target and not intersecting with the static obstruction volume, in order to highlight the motion assumptions that conform to the actual entry / exit path.

[0019] Optionally, the dynamic sparse voxel graph specifically involves: voxelizing multiple consecutive frames of sparse point cloud within the risk corridor at a fixed or segmented adaptive voxel scale to generate voxels with timestamps as nodes of the graph; the nodes record the voxel center, time index in the vehicle coordinate system, and occupancy count; the edges of the graph include at least cross-frame edges between adjacent frames, used to describe the spatial correspondence of potential dynamic targets in adjacent times;

[0020] The cross-frame edge connection is specifically as follows: between voxel nodes in adjacent frames, candidate edges are generated according to the principle of spatial proximity and temporal proximity. The edge records the displacement vector, displacement length and direction of the endpoint voxel center. The edge is subsequently determined by a motion voting field gating with consistent visibility to determine whether to retain it and its edge strength.

[0021] The visibility-consistent motion voting field gating is specifically as follows: using the visibility-consistent motion voting field output by S2 as the gating basis, for each candidate cross-frame connection, the following order is used for judgment and weighting: 1) The endpoint voxels are all located in the reachable space of potential dynamic targets and belong to the same risk corridor; 2) The angle between the displacement direction of the connection and the main direction of the risk corridor is calculated, and the angle does not exceed the direction tolerance recorded by the risk corridor; 3) Crossing detection is performed on the line segment connecting the centers of the endpoint voxels. If the line segment intersects with the static occlusion volume, the connection is rejected; 4) The visibility-consistent motion voting field values ​​in the direction of the connection are queried at the midpoint and two endpoints of the line segment, and the minimum value is used as the gating value of the connection. If it is lower than the gating threshold, it is rejected; 5) For the connection that passes the gating, its connection strength is adjusted according to the voting uncertainty. The greater the voting uncertainty, the lower the connection strength; the cross-frame connection after the above gating and weighting is used to aggregate the micro-motion evidence chain along the risk corridor in S4.

[0022] Optionally, the step of aggregating cross-frame occupancy changes and minute displacements along the risk corridor to form a consistent micro-motion evidence chain is as follows: In each risk corridor, a voxel located in the reachable space of a potential dynamic target and undergoing cross-frame occupancy changes is selected as the starting node, and the evidence chain is generated recursively along the cross-frame connection according to the following rules: 1) Both voxels at both ends of the connection are located within the risk corridor; 2) The consistency between the displacement direction of the connection and the main direction of the risk corridor at that location satisfies the direction tolerance; 3) The connection line segment does not intersect with the static occlusion volume; 4) When a short invisible segment appears, an empty segment is allowed to be inserted in the order of adjacent frames only in the neighborhood of the occlusion entrance and exit, and the recursion can only continue after the pairwise occupancy change confirmation is satisfied on both sides of the line of sight visibility boundary; 5) When multiple feasible connections appear, the joint sorting of connection strength and occupancy consistency weight is prioritized, and the second-best branch is reserved for parallel recursion; the formed evidence chain records the voxel index and connection set in chronological order for subsequent aggregation;

[0023] The provision allowing short invisible segments at obstructed entrances and exits, but requiring confirmation through paired occupancy changes on both sides of the visibility boundary, specifically means: when an observation gap appears in the chain of evidence at an obstructed entrance or exit due to its location on the invisible side, the crossing confirmation is made only if the length of the short invisible segment does not exceed the upper limit, based on paired occupancy changes on both sides of the visibility boundary; after confirmation, the invisible segment is incorporated into the temporal continuity of the chain of evidence according to the correspondence of boundary normal alignment, and this rule cannot be used in areas without obstructed entrances or exits;

[0024] The output of the three-dimensional verification candidate and evidence strength is as follows: the evidence chains of micro-motions with consistent directions within the same risk corridor are aggregated according to the spatial and temporal overlap relationship to obtain the three-dimensional verification candidate; the three-dimensional center is determined by the weighted average of the voxel centers on the chain, the orientation is determined by the weighted synthesis of the displacement directions of the connecting edges on the chain, and the evidence strength is determined by the weighted sum of the connecting edge strengths on the chain according to the occlusion consistency weight; the three-dimensional verification candidate and evidence strength are used as inputs of S5 for the initial estimation of the restricted state.

[0025] Optionally, the constraint estimation of candidate center, velocity and orientation is specifically as follows: taking the three-dimensional verification candidate output by S4 as input, the following steps are taken and constrained in the corresponding risk corridor: (1) Center constraint: taking the voxel center on the evidence chain that constitutes the candidate as a sample, the initial value of the center is obtained by weighting the product of the edge strength and the occlusion consistency weight; if the initial value is located in the static occlusion volume or the impassable area, it is projected back to the reachable space of the potential dynamic target along the normal of the nearest risk corridor center line, and it is ensured that it does not intersect with the static occlusion volume; (2) Velocity constraint: taking the projection of the cross-frame edge in the evidence chain in the risk corridor direction as the velocity sample, and retaining the angle between the edge and the risk corridor direction not exceeding the direction tolerance. The edges are connected and the velocity vector is synthesized by weighting the edge strength and the occlusion consistency weight; the normal component of the velocity is truncated with the upper limit corresponding to the direction tolerance, and the final velocity is mainly projected in the direction of the risk corridor; (3) Orientation constraint: the unit direction of the velocity vector is used as the initial value of orientation, and fine-tuning within the ± direction tolerance range is allowed to adapt to the local curvature of the risk corridor; if the orientation causes the displacement line segment to intersect with the static occlusion volume as time progresses, it will retreat to the nearest direction that satisfies the non-intersection constraint; (4) Consistency gating: for the above combination of center, velocity and orientation, it is required that it is within the risk corridor and does not cross the static occlusion volume, and is consistent with the direction of the motion voting field with visibility consistency. Combinations below the gating threshold are eliminated;

[0026] The use of occlusion consistency weight in the weighting process is as follows: In the weighted synthesis of center, velocity, and orientation, the edge strength of each cross-frame connection in the evidence chain constituting the 3D verification candidate is multiplied by the occlusion consistency weight of its spatial unit as the sample weight. Samples with higher weights contribute more to the estimation results. When the same candidate involves multiple risk corridor segments, only the orientation tolerance and occlusion consistency weight recorded in the risk corridor segment to which the candidate belongs are used for weighting to avoid cross-segment interference.

[0027] Optionally, the motion voting field based on visibility consistency extends along the risk corridor direction, and the extension terminates when it reaches the line-of-sight visibility boundary and there is a static occlusion volume in front. Specifically, it involves: determining the main direction and velocity magnitude within the current risk corridor based on the initial estimate of state S5; selecting an extension direction within the direction tolerance range along the main direction; advancing near the center line of the risk corridor by multiplying the adjacent frame time interval by the velocity magnitude as a step size, and querying the direction consistency value of the motion voting field with visibility consistency on the advancing line segment, requiring that the value is not lower than the gate threshold; continuously constraining the line segment to be completely located in the risk corridor and not intersecting with the static occlusion volume during the advancement process; when the ray along the extension direction first reaches the line-of-sight visibility boundary and there is a static occlusion volume in front of it, the extension terminates at the line-of-sight visibility boundary and is marked as the extension endpoint; otherwise, it continues to extend under the above gate and constraint conditions until the next observation segment appears or the termination condition is met.

[0028] The process of completing the reconstruction when a new micro-motion evidence chain appears in a subsequent frame and is consistent with the original risk corridor is as follows: continuously monitor the micro-motion evidence chain output by S4 in the neighborhood of the extended endpoint and dock it according to the following consistency rules: (1) Spatial consistency: the distance between the starting voxel center of the evidence chain and the extended endpoint does not exceed the bandwidth of the risk corridor at that location, and both are located in the reachable space of potential dynamic targets; (2) Directional consistency: the angle between the displacement direction of the connecting edge of the evidence chain and the main direction of the risk corridor at that location does not exceed the directional tolerance, and is in the same direction as the extension direction; (3) Visibility consistency support: the minimum value of the line segment connecting the extended endpoint and the starting point of the evidence chain in the visibility consistency motion voting field is not lower than the gate threshold; (4) Occlusion geometric consistency: the above connecting line segment does not cross the static occlusion volume and conforms to the lateral relationship of the line of sight visibility boundary; when all the rules are met, the evidence chain is connected to the trajectory and the extended segment is replaced with the observation data of the evidence chain to complete the reconstruction; when there are multiple candidate evidence chains, the one with the highest evidence strength is docked first, and the unselected candidate does not change the current trajectory;

[0029] The beneficial effects of this invention are:

[0030] This scheme unifies the encoding of sparse point clouds, road topology, passability and pedestrian crossing semantics, and static occlusion volume in the vehicle coordinate system. It uses ray-marking to define the visibility boundary and occlusion entrances / exits, and combines multi-agent contextual reasoning to generate a visibility-consistent motion voting field. At entrances / exits and intersections, it enhances the orientation according to the passability boundary normal, extracting risk corridors with weighted directional tolerance and occlusion consistency. Unlike approaches that rely solely on free space inference or heuristic inflation, this design explicitly defines the actual path for potential targets to enter the visible side from the invisible side using closed occlusion volume and lateral relationships, ensuring that the reachable direction aligns with road geometry and reducing motion assumptions inconsistent with the scene structure.

[0031] Within the risk corridor, a dynamic sparse voxel map is constructed by voxelizing multi-frame sparse point clouds. Cross-frame edges are gated using a voting field, retaining only edges consistent with the main direction of the corridor and not crossing static occlusion volumes. When the corridor crosses multiple lanes, gating is performed independently according to the segmented main direction and direction tolerance to avoid false connections between corridors. Cross-frame occupancy changes and small displacements are aggregated along the corridor to form a micro-motion evidence chain. Short invisible segments are allowed only in the neighborhood of occlusion entrances and exits, and confirmed by paired occupancy changes on both sides of the visibility boundary. Compared to schemes that rely solely on detection results or global data association, this evidence and geometrically consistent aggregation mechanism can maintain spatiotemporal consistency at occupancy transitions, suppressing false cross-lane associations and spurious connections on invisible sides.

[0032] After obtaining 3D verification candidates based on the evidence chain, a constrained state estimation method is used to constrain the center, velocity, and orientation by the risk corridor direction tolerance and occlusion consistency weights. The velocity is mainly projected onto the corridor direction, and if the center falls into a static occlusion volume or an impassable area, it is projected back to the potential reachable space. In the event of short-term invisibility or a sudden decrease in points, the system extends along the corridor direction according to the voting field, and completes the reconstruction when the subsequent evidence chain is consistent with the original corridor. The system outputs the target and its associated risk corridor, and uses the intersection of the corridor with the vehicle's travel path or a passable area to form a potential conflict zone for early warning. Compared with processes lacking visibility modeling, this link constrains and continuously verifies the entry channel around the occlusion entrance and exit, which helps maintain trajectory continuity during the invisible-to-visible transition and provides actionable scenario-based risk warnings to the decision-making level while meeting visibility consistency requirements. Attached Figure Description

[0033] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention, but do not constitute a limitation thereof; in the drawings:

[0034] Figure 1 This is a flowchart of a deep learning-based method for identifying and tracking occluded dynamic targets proposed in this invention.

[0035] Figure 2This is a schematic diagram of visibility-consistent motion corridor occlusion extension and reproduction in a deep learning-based method for identifying and tracking occluded dynamic targets proposed in this invention. Detailed Implementation

[0036] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams illustrating only the basic structure of the invention and therefore only show the components relevant to the invention.

[0037] refer to Figure 1 and Figure 2 A deep learning-based method for identifying and tracking occluded dynamic targets, comprising:

[0038] S1. On roads with rain, snow, fog, night, and large obstructions, using the vehicle's pose as a reference, sparse point clouds are uniformly encoded with road topology, passable areas, pedestrian crossing areas, and static obstruction volumes. Ray testing is performed on the obstruction volumes, the visibility boundaries and obstruction entrances and exits are marked, and the visible / invisible sides are distinguished to generate scene element maps to limit the reachable space of potential dynamic targets.

[0039] S2. Based on the scene element map of S1, the potential dynamic target is explicitly modeled as an agent using a multi-agent contextual reasoning network. Combining road topology, pedestrian crossing and vehicle flow interaction patterns and occlusion geometry, the existence probability, movement direction and occlusion degree are inferred. A movement voting field with consistent visibility is output. At the occlusion entrances and exits and intersections, the normal direction of the passable boundary is enhanced. A risk corridor is generated in the voting enhancement zone. The risk corridor carries the direction tolerance and occlusion consistency weight.

[0040] S3. Based on the risk corridor of S2, multi-frame sparse point cloud voxels are formed in the corridor to form a dynamic sparse voxel map. Cross-frame edges are gated by a motion voting field with consistent visibility. Only voxels that are in the same direction as the risk corridor and do not cross static occlusion volumes are connected. The edge strength is adjusted according to the voting uncertainty. When the corridor crosses multiple lanes, the edges are independently gated with tolerance in each direction to avoid misconnection between corridors.

[0041] S4. On the dynamic sparse voxel map of S3, aggregate cross-frame occupancy changes and small displacements along the risk corridor to form a chain of micro-motion evidence with consistent direction. Suppress segments that cross static occlusion volumes or deviate from the direction of the risk corridor. Allow short invisible segments for occlusion entrances and exits, but pairwise occupancy changes must appear on both sides of the visibility boundary for confirmation. Output three-dimensional verification candidates and evidence strength.

[0042] S5. Based on the three-dimensional verification candidates of S4, combined with the directional tolerance and occlusion consistency weight of S2, the candidate center, velocity and orientation are constrained and estimated to obtain the initial state estimate and uncertainty interval. The constraints include: the center must not fall into the static occlusion volume or the impassable area, the velocity is mainly projected in the risk corridor direction, and the occlusion consistency weight is used to participate in the weighting to suppress unphysical movements.

[0043] S6. Based on the initial state estimate in S5, a time-continuous trajectory is generated within the risk corridor. When it is temporarily invisible or the number of points decreases sharply, it extends along the direction of the risk corridor based on the motion voting field with consistent visibility. The extension terminates when it reaches the visibility boundary and there is a static occlusion volume in front. When a new micro-motion evidence chain appears in subsequent frames and is consistent with the original risk corridor, the reproduction is completed, reducing trajectory interruption.

[0044] S7. Based on the trajectory of S6 and the three-dimensional verification candidates of S4, output the target position, speed, orientation and the risk corridor to which it belongs to the vehicle decision-making.

[0045] In this embodiment, the specific implementation of step S1 is as follows:

[0046] In step S1, firstly, using the vehicle's pose as a reference, all elements of a frame (time index k) are aligned to the vehicle's coordinate system. Let the vehicle's pose in this frame be a rigid body transformation H. k =(R k ,t k ), where R k ∈SO(3) is a rotation matrix. This is a translation vector. It represents the sparse point cloud in the sensor coordinate system. Transformed into the vehicle coordinate system points

[0047] p = R k p S +t k

[0048] And recorded as the viewpoint of the vehicle This is for subsequent ray testing. This step also unifies the data on road topology, passable areas, pedestrian crossing areas, and static occlusion volumes into the same vehicle coordinate system.

[0049] To achieve unified coding, a unified spatial index is used to align features, with three-dimensional voxels being preferred (though this can also degenerate into a two-dimensional grid). Let the voxel origin be... Voxel side length δ>0, voxel index set The voxel index and voxel center of any point p are defined as follows:

[0050]

[0051] where i=(ix i y i z )∈Ω is the voxel index. For the voxel center, 1 = (1,1,1). This is rounded down to the nearest integer. A two-dimensional mesh can be considered as a fixed i. z The slices are defined consistently.

[0052] Semantic labels and adjacency relationships are recorded on the same voxel unit to complete unified encoding. A semantic label mapping is defined. Its components are in For road topology identification (discrete labeling), This is a passable attribute. For pedestrians crossing the road, For occlusion properties. Define the set of adjacency relationships. in For lane connection relationships, As for the boundary relationship of the passable area, This refers to the contact relationships of statically occluded volumes. It includes contact relationships between adjacent voxel pairs at the boundary. Record local normal u ij =(c j -c i ) / ||c j -c i ||2, where ||·||2 is the Euclidean norm.

[0053] To determine obstruction and calculate visibility under conditions of rain, snow, fog, nighttime, and large obstructions, a meteorological visibility radius ρ is introduced. k >0 limits the effective viewing distance; use a static occlusion volume voxel set. and its indicator function χ O : This indicates occupancy (1 for occlusion, 0 for freedom). For each voxel center c i Rays emitted from the viewpoint i (t)=s+t(c i -s)(t∈[0,1]), calculate the first intersection parameters:

[0054] τ i =inf{t∈[0,1]|χ O (r i (t))=1}

[0055] If τ i It does not exist and ||c i -s||2≤ρ k If a voxel is considered visible, it is determined to be on the visible side; otherwise, it is considered invisible. Let the set of visible voxels be denoted as . The invisible voxel set is

[0056] Based on the visibility results, the line-of-sight visibility boundaries and occlusion entrances / exits are extracted. Let N(i) be the voxel neighborhood (e.g., 6-neighborhood), and the line-of-sight visibility boundary set is defined as:

[0057]

[0058] For each i∈Γ, estimate the occlusion outward normal. (Unit vector), and with Distinguish between visible and invisible sides. Denote the set of voxels representing the passable region as... The set of voxels of the area crossed by pedestrians is denoted as The set of blocked entrances and exits is defined as follows:

[0059]

[0060] Within a unified coding space, the reachability space of a potential dynamic target is defined, denoted as . It is obtained by subtracting the static obstruction volume from the passable area and the area crossed by pedestrians:

[0061]

[0062] The final scene element map is denoted as... The scene element map achieves unified encoding of sparse point cloud and road topology, passable area, pedestrian crossing area and static occlusion volume. It also marks the line-of-sight visibility boundary and occlusion entrances and exits through ray testing and distinguishes between visible and invisible sides, thereby limiting the reachable space of potential dynamic targets and providing a consistent spatial and semantic basis for subsequent steps.

[0063] In this embodiment, the specific implementation of step S2 is as follows:

[0064] In step S2, the scene element map obtained in the previous step... Using this as input, a multi-agent contextual reasoning network is constructed within the vehicle's coordinate system. Firstly... Extracting the node set of the contextual reasoning graph Includes lane segment nodes (by) and (Aggregated), pedestrian crossing area nodes (from) (Aggregated), occluded entrance / exit nodes (i.e., ε), and passable boundary nodes (obtained by aggregation). (Aggregate to obtain); simultaneously construct an edge set Including the connection relationships of road topology (sourced from) The conflict relationship between pedestrian crossing areas and lanes at the conflict point (represented by directed edges to indicate yielding / priority order), and the visible / invisible side relationship of the visibility boundary (determined by Γ and Γ). (Derived). On this graph, the network at the occlusion entrance / exit ε and the intersection (the set of nodes determined by topology and conflict relationships, denoted as...) Explicitly generate surrogate hypothesis sets Each agent The state is in Candidates for center location (constrained within the reachable space of potential dynamic targets) Inside), For the candidate (unit direction vector), π a ∈[0,1] represents the probability of existence, ζ a ∈[0,1] represents the degree of occlusion. Network message passing and constraint reasoning simultaneously encode traffic flow interaction patterns: following relationships within the same lane, yielding relationships at merging / exit points, and the crossing priority relationships between lanes and pedestrian crossing areas at conflict points. The directed constraints in the system restrict the relative direction and order of passage between agents, making the agent set consistent with physical reachability.

[0065] To generate a motion voting field with consistent visibility, targeting the voxel center c i (i∈Ω) and direction The orientation of each agent is integrated with spatial support, occlusion consistency, and boundary orientation enhancement to define a voting value V(c). i (d). This quantity is the core calculation point in this step, used to simultaneously ensure "visibility consistency" and "physical accessibility consistency," and directly serves the subsequent formation and targeted enhancement of risk corridors:

[0066]

[0067] Where π a The probability of an agent existing; As the direction kernel, the metric d is... Consistency, σ θ >0 represents the directional tolerance (in radians); For the space kernel, measure c i relatively Spatial support, σ x >0 represents the spatial bandwidth; ω occ (c i )∈[0,1] represents the occlusion consistency weight, based on the visible / invisible side of Γ and the occlusion geometry pair c. i empowerment; g bnd (c i d) is the directional enhancement factor along the traversable boundary normal; For the potential dynamic target reachability space indicator function (voxel center at) (Take 1 if it's internal, otherwise take 0). The kernel function can be Gaussian or Von Mises-Fisher type, and its values ​​are dimensionless, ensuring that V(·) is a dimensionless voting strength. Thus, only when both conditions are met (by...) can the kernel function achieve its maximum value. Consistent in direction (by κ), consistent in space (by η), and consistent with visibility (by ω) occ And through boundary normal orientation enhancement (by g) bnd Only by adopting the motion hypothesis can significant support be obtained on V.

[0068] Directional augmentation follows the principle of directional augmentation along the normal of the traversable boundary: at the intersection of the blocked entrance / exit ε. Within the neighborhood, the motion component pointing towards the visible side or the conflict line normal is enhanced based on the local normal of the traversable area boundary. Let it be... In c i The unit normal obtained by neighborhood aggregation is (Taking the normal direction from the invisible side to the visible side at the obstructed entrance / exit, and taking the bidirectional direction according to the conflict line at the intersection), the orientation enhancement factor is defined as follows:

[0069]

[0070] Where α∈[0,1] is the enhancement coefficient, and · is the inner product operation, ensuring that the directional component "from the invisible side to the visible side" is preferentially enhanced at the obstructed entrance / exit, and that symmetrical enhancement is performed on both sides of the conflict line at the intersection. Combined with ω occ (·) Consistency constraints on the visible / invisible side, as described above g bnd (·) Only in It operates on voting components that do not intersect with static occlusion volumes, thereby highlighting the assumptions that conform to actual entry and exit paths and converging movements.

[0071] Based on the obtained consistent visibility of the motion voting field V(c) i d) Create a voting enhancement zone at the entrance / exit and intersection, and along the main voting direction. The inner ridge line is extracted as the centerline of the risk corridor; for each segment of the risk corridor, the directional tolerance σ is recorded. θ Corridor-level convergence with occlusion consistency weight (e.g., for ω along the line) occ (·) weighted average), and these attributes are stored along with the corridor. The resulting output shows a consistent visibility of the motion voting field not only spatially with Consistent with Γ in terms of visibility, and also through g at the obstruction of entrances and intersections. bnd(·) "Directional enhancement according to the normal direction of the passable boundary" is realized, and finally a risk corridor carrying directional tolerance and occlusion consistency weights is generated within the enhancement band, which can be directly referenced and gating in subsequent steps S3–S7.

[0072] In this embodiment, the specific implementation of step S3 is as follows:

[0073] In step S3, the scene element map obtained in the previous step... Movement voting field V(c) consistent with visibility i Taking (d) as input, a dynamic sparse voxel graph is constructed within each risk corridor output from step S2. The set of risk corridors is denoted as... Each of the corridors Record the principal direction field near its centerline (Unit vector), Direction tolerance (Radians) and corridor-level convergence weights for occlusion consistency. For consecutive frames of sparse point clouds, the voxel parameters δ defined in step S1 and the index function i(·) are used in each... Voxelization is performed internally, generating voxel nodes with time indexes. Each node records the voxel center. Time Index And occupancy count. Between adjacent frames k and k+1, according to the principle of spatial proximity, in the same Candidate cross-frame edges are generated internally, and the edge e connects to the voxel index (i). k ,j k+1 ), and record the displacement vector. Displacement length ||v e ||2 and unit direction When the corridor crosses multiple lanes, combined The corridor is managed in segments, with each segment using its own σ. θ (·) Independent gating to avoid incorrect connections across segments.

[0074] Candidate cross-frame edges need to be gated by a visibility-consistent motion voting field, retaining only edges that align with the risk corridor direction and do not traverse static occlusion volumes. Let the indicator function for the reachability space of potential dynamic targets be... Let the corridor member indicator function be: Assume the non-crossing detection indicator is 1. occ (e)∈{0,1} (take 1 when the connecting line segment does not intersect with the static occlusion volume, otherwise take 0). Let the midpoint of the connecting edge be... Set the angle consistency indicator to 1. ang (e)∈{0,1}(when Take 1 at the time.

[0075] Define the edge gate value as follows:

[0076]

[0077] Where γ(e)∈[0,1] is a dimensionless gating value; V(·,·)∈[0,1] is a visibility-consistent motion voting field; The main direction of the corridor; For corridor direction tolerance; Ensure that the endpoint is located in the reachable space of the potential dynamic target; Ensure that the endpoints belong to the same risk corridor; 1 ang (e) Ensure the direction of the connecting edge is consistent with the main direction of the corridor; 1 occ (e) Ensure that the connecting line segments do not intersect with the static occlusion volume. Using a threshold τ... V Determine whether to retain the edge ∈(0,1] (γ(e)≥τ) V (Then retain), thereby applying the visibility consistency assumption of S2 and the reachability and occlusion geometric constraints of S1 together in the same gating, directly suppressing cross-occlusion and out-of-bounds false connections.

[0078] For gating edges, to reflect the principle of "adjusting edge strength according to voting uncertainty," the strength is monotonically suppressed without increasing gating complexity. Let the voting uncertainty field be U(c,d)∈[0,1] (derived from the network output in step S2 or from the local concentration of V; a larger value indicates higher uncertainty), and let the edge uncertainty be... Define the edge strength as w e =γ(e)(1-u e ), where w e ∈[0,1] represents a dimensionless weight. Ultimately, in each risk corridor... The resulting dynamic sparse voxel map connects only voxels aligned with the risk corridor direction and not traversing static occlusion volumes, and its intensity is adaptively reduced based on voting uncertainty. When the corridor crosses multiple lanes, each segment independently uses the corresponding σ. θ (·) Gating is implemented to prevent accidental connections between corridors. The above structure provides a consistent, visible, and geometrically consistent graphical basis for aggregating the micro-motion evidence chain along the risk corridor in step S4.

[0079] In this embodiment, the specific implementation of step S4 is as follows:

[0080] In step S4, based on step S3, for each risk corridor set The dynamic sparse voxel map obtained within the field along the corresponding risk corridor principal direction. Orientation tolerance By aggregating cross-frame occupancy changes and minute displacements, a chain of evidence of micro-motions with consistent orientation is formed. Specifically, this is achieved within the reachable space of potential dynamic targets. Inside, and belonging to the risk corridor Among the voxel nodes, the voxel that undergoes cross-frame occupancy change is selected as the starting node (the occupancy count recorded by the node is denoted as...). Where i∈Ω is the voxel index. (For time index); the recursion only extends along the cross-frame connection e that has been gated in step S3 (the connection endpoint voxel center c). i With c j The midpoint is The unit direction of displacement is Strength is w e ∈[0,1]), and follows the rule: endpoints are of the same origin. direction and Consistent and not exceeding σ θ (·) The connecting line segments do not intersect with the static occlusion volume. In this way, segments that deviate from or cross the static occlusion volume are directly suppressed.

[0081] To address the special case where a "short invisible segment" is allowed in the neighborhood of an obstructed entrance / exit, a method of "confirming the paired occupancy changes on both sides of the line-of-sight visibility boundary" is introduced to bridge the observation gap. Let the set of boundary voxel indices be Γ, and the set of obstructed entrances / exits be... The voxel neighborhood is N(·), and the sets of visible / invisible voxels are respectively The passable and pedestrian-crossing voxel sets are respectively When the chain of evidence is in i b When a gap occurs at position ∈ε, from frame k to frame k+Δk (Δk is the gap frame length), the gap can only be crossed if the confirmation gate is 1 and the following condition is met:

[0082] Φ pair (i b ,k,Δk)=1 ε (i b )1 len (Δk)1{Δn p,k ≤-τ o}1{Δn q,k ≥τ o}

[0083] Among them 1 ε (i b )∈{0,1} indicates an obstructed entrance / exit; 1 len (Δk)∈{0,1} is a length indicator (when Δk≤N) max Take 1 at time. (The upper limit of the short invisible segment); Visible neighboring voxels, Invisible neighboring voxels; Δn p,k =n p,k+1 -n p,k With Δn q,k =n q,k+1 -n q,k The change in occupancy count between adjacent frames; τ o >0 represents the occupancy change threshold. This gating requires that within the same neighborhood at any given time, the visible side experiences a decrease in occupancy while the invisible side experiences an increase in occupancy (or is set symmetrically as needed), and the gap length is limited, thus ensuring that the "invisible segment" is only accepted under the actual occupancy geometry of entry from the visible side to the invisible side. This rule is not allowed except in occupancy entrance / exit neighborhoods.

[0084] When multiple feasible edges appear during the recursive process, the one with the higher joint strength-occlusion consistency is selected first (strength comes from w). e The occlusion consistency is determined by ω given in step S2. occ (c)∈[0,1] at the midpoint m of the connecting edge e The values ​​at each point are considered, and a small number of suboptimal branches are retained for parallel recursion to cover multi-target or multi-path scenarios. After recursion, micro-motion evidence chains within the same risk corridor that are consistent in direction, continuous in time, and satisfy the above rules are aggregated according to spatial and temporal overlap to form three-dimensional verification candidates. The three-dimensional center of the three-dimensional verification candidate is obtained by weighted averaging of the voxel centers on the chain, and the orientation is obtained by weighted synthesis of the displacement unit directions on the chain; the evidence strength is weighted and accumulated according to the edge strength and occlusion consistency weights, defined as:

[0085]

[0086] in For the chain of evidence Strength of evidence, w e ∈[0,1] represents the edge strength obtained in step S3, ω occ (m e The occlusion consistency weight is defined as ∈ [0,1]. The final output is each 3D validation candidate and its... As input for the initial estimation of the constrained state in step S5.

[0087] In this embodiment, the specific implementation of step S5 is as follows:

[0088] In step S5, the three-dimensional verification candidate obtained in the previous step is used as input, and a preliminary assessment of the confined state is performed within the corresponding risk corridor. Let a certain three-dimensional verification candidate be denoted as... Its associated risk corridor is The candidate consists of the micro-motion evidence chain aggregated in step S4, and its set of voxel node indices is denoted as . The set of cross-frame edges included is Each cross-frame connection Connect the two voxel centers The displacement vector is Midpoint The edge strength is w e ∈[0,1], direction consistency indicator is 1 ang (e)∈{0,1} (all from step S3), the occlusion consistency weight is ω occ (·)∈[0,1](from step S2). The main direction field of the corridor is denoted as... Orientation tolerance is The static occlusion volume is determined by the indicator function χ. O : Given (1 for occlusion, 0 for freedom), the reachability space of a potential dynamic target is indicated as follows: Corridor members are instructed as follows The visibility-consistent motion voting field is V(c,d)∈[0,1] (from step S2).

[0089] First, center constraint estimation is performed. Using voxel centers on the evidence chain as samples, initial center values ​​are obtained by weighting the product of edge strength and occlusion consistency weights, ensuring that samples come only from the corridor segments to which the candidates belong. Node weights are defined for this purpose. in For voxel indexing, Let it be the center. The initial value of the center is calculated as follows:

[0090]

[0091] when or At that time, the normal direction along the center line of the nearest risk corridor will Projection Back And satisfy And ensure that the projection result does not intersect with the static occlusion volume, to obtain the center estimate c. ★ .

[0092] Next, velocity constraint estimation is performed. The projection of cross-frame edges onto the risk corridor direction is used as the velocity sample. Edges satisfying the direction tolerance are retained. The velocity is weighted and synthesized using edge strength and occlusion consistency weights, while suppressing the normal component. Assuming the time interval between adjacent frames is Δt>0, the constrained velocity estimate is:

[0093]

[0094] in The displacement is a scalar projection (dimension: length) along the main direction of the corridor, normalized to velocity (dimension: length / time) by Δt, and synthesized as a velocity along the corridor. The velocity vector. Its normal component is... The corresponding upper limit is truncated so that the final velocity is mainly projected in the direction of the risk corridor.

[0095] Orientation constraint estimation is performed again. The unit direction of the velocity vector is used as the initial orientation value, denoted as... And allow in Fine-tuning is performed within the range to adapt to the local curvature of the corridor. When along... The advancing displacement segment intersects with the static shading volume (by χ). O During detection, backtrack to the nearest direction that satisfies the disjoint constraint. Regarding consistency gating, for... Requirements must be met simultaneously The advancing line segment does not cross the static occlusion volume, and (τ V (∈(0,1] is the gating threshold). When the same 3D verification candidate spans multiple risk corridor segments, only the directional tolerance σ recorded in its respective segment is used. θ (·) and occlusion consistency weight ω occ (·) Participate in weighting and gating to avoid cross-segment interference. Final output state initial estimation. and its uncertainty range.

[0096] In this embodiment, the specific implementation of step S6 is as follows:

[0097] In step S6, the initial estimate of the constrained state from the previous step is... Starting from the relevant risk corridor The trajectory is generated continuously over time and extends along the risk corridor along a motion voting field with consistent visibility when it is temporarily invisible or when the number of points decreases sharply, until the termination condition is met. For the center estimate, For constrained velocity estimation (dimensions: length / time), For orientation estimation (unit direction vector), Carrying the main direction field Orientation tolerance The visibility-consistent motion voting field is denoted as V(c,d)∈[0,1], and the potential dynamic target reachability space indicator function is denoted as... The corridor member indicator function is denoted as The static occlusion volume is determined by the indicator function χ. O (c)∈{0,1} is given (1 represents occlusion), and the set of line-of-sight visibility boundaries is Γ. To ensure dimensional consistency, the stretching step size is the product of the time interval Δt>0 between adjacent frames and the velocity magnitude. (Dimension is length), extension direction is taken as the main direction of the corridor. And allow in Fine-tuning within the range to adapt to local curvature.

[0098] The extension employs a joint gating system consistent with both the voting field and geometry, proceeding only under the conditions of "reachable—within the corridor—without traversing obstructions—and sufficient voting support." The core of this step is the extension gating value, which applies to the extension segment. Sampling and combining the indicator function with the non-crossing detection: Let the non-crossing detection indicator be 1. occ (l)∈{0,1} if and only if line segment l does not intersect with the static occlusion volume (i.e., points along l satisfy χ). O When (·) = 0), the value is 1, and the extended gate value is defined as:

[0099]

[0100] Among them G ext ∈[0,1] is a dimensionless gate value, and min{·} ensures that the minimum value of the voting support consistent along the direction of the extended line segment is not lower than the gate threshold. and Ensure that the endpoints are within reachable space and belong to the same risk corridor, 1 occ (l) Ensure that the extended line segment does not cross the static occlusion volume. Using a threshold τ... V ∈(0,1] Decision progression: when G ext ≥τ V At that time, generate extended endpoints. The extended segment will be incorporated into the trajectory; otherwise, the extension or shrinkage will be rejected and retried.

[0101] The termination condition for the extension follows the rule of "reaching the line-of-sight visibility boundary and having a static occlusion volume in front". Specifically, if there exists τ Γ ∈(0,1] makes And for any sufficiently small ε>0, we have Then terminate the extension at the visibility boundary and mark the extension endpoint; otherwise, satisfy G. ext ≥τ V Under the premise of geometric constraints, it continues to extend until the next observation segment appears or the termination condition is met.

[0102] To achieve reproduction, the micro-motion evidence chain output in step S4 is continuously monitored in the extended endpoint neighborhood, and the docking is completed according to the consistency rule:

[0103] First, spatial consistency requires that the center of the starting voxel of the chain of evidence be consistent with c. ext The distance between them does not exceed the corridor bandwidth at that location, and both are located in and Inside;

[0104] Secondly, directional consistency requires that the directions of displacement of the connecting edges of the evidence chain be consistent with... The included angle does not exceed And in the same direction as the extension;

[0105] Third, consistent visibility support requires connection c ext The minimum value of the line segment connecting the starting point of the evidence chain on V(·,·) is not less than τ. V ;

[0106] Fourth, the occlusion geometry consistency requirement stipulates that the connecting line segment does not cross the static occlusion volume and satisfies the lateral relationship of the line-of-sight visibility boundary. When all rules are satisfied, the evidence chain is linked into the trajectory, and the extended segment is replaced with the observation data of the evidence chain to complete the reconstruction; when there are multiple candidate evidence chains, the one with the highest evidence strength given in step S4 is prioritized for connection. The above extension-termination-reconstruction mechanism ensures that the trajectory remains robustly connected along the risk corridor even when it is not visible for a short time, and is always subject to V(·,·), Γ and χ. O Joint constraints of (·).

[0107] In this embodiment, the specific implementation of step S7 is as follows:

[0108] In step S7, based on step S6, in the corresponding risk corridor The obtained continuous time trajectory and the 3D verification candidates output in step S4 are used to form a robust output quadruple for vehicle decision-making for each target. Let the latest trajectory state of the target be... in Location (from the latest center of S6), For velocity (constrained velocity estimate from S6, in units of length / time), Let the orientation be (unit direction vector); let the 3D verification candidate matching this trajectory in the spatial and temporal neighborhood be... (From S4), its three-dimensional center and orientation are respectively denoted as... and Strength of evidence (Weighted cumulative strength of the candidate evidence chain). To enhance the stability of the output under occlusion conditions, a position fusion driven by the visibility-consistent motion voting field V(c,d)∈[0,1] (from S2) and the evidence strength is adopted. The dimensionless fusion coefficient λ∈(0,1) is defined as a monotonically increasing function. The output of, where For logical functions, parameters S is the calibration coefficient. max >0 represents the saturation value for the strength of evidence. Therefore, the position output is:

[0109]

[0110] in This is the output position (measured in length).

[0111] The above formula uses the voting support of the trajectory. Strength of evidence for candidates Joint control λ, biased towards c when trajectory reliability is high. trk High-evidence candidate correction is applied in the early stages of reproduction or near the occlusion boundary. Ensure consistency with visibility and physical reach. Velocity output is taken as... To maintain a restricted velocity estimate consistent with the risk corridor direction; output-oriented priority. when and When the threshold is not lower than the threshold, use Fine-tuning is performed to adapt to local curvature. Finally, a quadruple is output to the vehicle's decision-making system. And apply a consistency check: require p out With v out The corresponding propulsion segment lies entirely within the reachable space of the potential dynamic target (as indicated by the indicator function). (constraints) and not related to the static occlusion volume (as indicated by the indicator function χ). O (·) gives the intersection, and The output must not fall below the gating threshold; if the threshold is violated, the output will fall back to the state that most recently satisfied the constraint or maintain the previous valid output, thereby ensuring the reliability and safety of the output in rain, snow, fog, nighttime, and large obstruction scenarios.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying and tracking occluded dynamic targets based on deep learning, characterized in that, include: S1. On roads with rain, snow, fog, night, and large obstructions, using the vehicle's pose as a reference, sparse point clouds are uniformly encoded with road topology, passable areas, pedestrian crossing areas, and static obstruction volumes. Ray testing is performed on the obstruction volumes, the visibility boundaries and obstruction entrances and exits are marked, and the visible / invisible sides are distinguished to generate scene element maps to limit the reachable space of potential dynamic targets. S2. Based on the scene element map of S1, the potential dynamic target is explicitly modeled as an agent using a multi-agent contextual reasoning network. Combining road topology, pedestrian crossing and vehicle flow interaction patterns and occlusion geometry, the existence probability, movement direction and occlusion degree are inferred. A movement voting field with consistent visibility is output. At the occlusion entrances and exits and intersections, the normal direction of the passable boundary is enhanced. A risk corridor is generated in the voting enhancement zone. The risk corridor carries the direction tolerance and occlusion consistency weight. S3. Based on the risk corridor of S2, multi-frame sparse point cloud voxels are formed in the corridor to form a dynamic sparse voxel map. Cross-frame edges are gated by a motion voting field with consistent visibility. Only voxels that are in the same direction as the risk corridor and do not cross static occlusion volumes are connected. The edge strength is adjusted according to the voting uncertainty. When the corridor crosses multiple lanes, the edges are independently gated with tolerance in each direction to avoid misconnection between corridors. S4. On the dynamic sparse voxel map of S3, aggregate cross-frame occupancy changes and small displacements along the risk corridor to form a chain of micro-motion evidence with consistent direction. Suppress segments that cross static occlusion volumes or deviate from the direction of the risk corridor. Allow short invisible segments for occlusion entrances and exits, but pairwise occupancy changes must appear on both sides of the visibility boundary for confirmation. Output three-dimensional verification candidates and evidence strength. S5. Based on the three-dimensional verification candidates of S4, combined with the directional tolerance and occlusion consistency weight of S2, the candidate center, velocity and orientation are constrained and estimated to obtain the initial state estimate and uncertainty interval. The constraints include: the center must not fall into the static occlusion volume or the impassable area, the velocity is mainly projected in the risk corridor direction, and the occlusion consistency weight is used to participate in the weighting to suppress unphysical movements. S6. Based on the initial state estimate in S5, a time-continuous trajectory is generated within the risk corridor. When it is temporarily invisible or the number of points decreases sharply, it extends along the direction of the risk corridor based on the motion voting field with consistent visibility. The extension terminates when it reaches the visibility boundary and there is a static occlusion volume in front. When a new micro-motion evidence chain appears in subsequent frames and is consistent with the original risk corridor, the reproduction is completed. S7. Based on the trajectory of S6 and the three-dimensional verification candidates of S4, output the target position, speed, orientation and the risk corridor to which it belongs to the vehicle decision-making.

2. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The specific method of uniformly encoding sparse point clouds with road topology, passable areas, pedestrian crossing areas and static occlusion volumes is as follows: under the vehicle coordinate system defined by the vehicle pose, a unified spatial index is used to align the above elements to the same reference frame, and semantic labels and adjacency relationships are recorded on the same spatial unit. The spatial index is a two-dimensional grid or a three-dimensional voxel; the semantic tags include road topology identifiers, passability attributes, pedestrian crossing attributes, and occlusion attributes; the adjacency relationships include lane connection relationships, contact relationships between passable area boundaries and static occlusion volumes, to ensure that each element can be directly used for occlusion determination and visibility calculation in rain, snow, fog, nighttime, and large occlusion scenarios.

3. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The multi-agent contextual reasoning network is specifically as follows: taking the scene element graph obtained in S1 and the reachability space of potential dynamic targets as input, a graph composed of lane segments, pedestrian crossing areas, occluded entrances and exits, and passable boundary nodes is constructed in the vehicle coordinate system; the connection relationship of the road topology, the conflict relationship between the pedestrian crossing area and the lane, and the visible / invisible side relationship of the line-of-sight visibility boundary are used as the edges of the graph; potential dynamic targets are explicitly modeled as agents on the above graph, that is, agent hypotheses are generated at the nodes of occluded entrances and exits and intersections, and the agent's state includes center position candidate, orientation candidate, existence probability, and occlusion degree; the network limits the allowed movement direction of the agents based on the road topology and passable areas, and limits the visibility consistency of the agents based on the occlusion geometry, thereby outputting an agent set consistent with physical reachability for subsequent movement voting; The specific traffic flow interaction rules are as follows: Within the scene element graph, the following relationship along the lane direction within the same lane, the yielding relationship between adjacent lanes at merging / exit points, and the crossing priority relationship between lanes and pedestrian crossing areas at conflict points are expressed in a regularized manner. The above relationships are mapped as directed constraints of the graph, which are used to restrict the relative direction and passability order between agents in the multi-agent contextual reasoning network. The directional enhancement based on the normal of the accessible boundary specifically involves: within the neighborhood of the obstructed entrance / exit and the intersection, directional gain is applied to the voting components pointing towards the visible side based on the local normal of the accessible boundary; for the obstructed entrance / exit, the normal pointing from the invisible side to the visible side is taken as the priority direction; for the intersection, the bidirectional normal corresponding to the conflict line is taken as the priority direction; the directional enhancement only applies to the voting components located in the reachable space of the potential dynamic target and not intersecting with the static obstruction volume, in order to highlight the motion assumptions that conform to the actual entry / exit path.

4. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The dynamic sparse voxel graph is specifically described as follows: within the risk corridor, multiple consecutive frames of sparse point clouds are voxelized at a fixed or segmented adaptive voxel scale to generate voxels with timestamps as nodes of the graph; the nodes record the voxel center, time index in the vehicle coordinate system, and occupancy count; the edges of the graph include at least cross-frame edges between adjacent frames, used to describe the spatial correspondence of potential dynamic targets in adjacent times. The cross-frame edge connection is specifically as follows: between voxel nodes in adjacent frames, candidate edges are generated according to the principle of spatial proximity and temporal proximity. The edge records the displacement vector, displacement length and direction of the endpoint voxel center. The edge is subsequently determined by a motion voting field gating with consistent visibility to determine whether to retain it and its edge strength. The visibility-consistent motion voting field gating is specifically as follows: using the visibility-consistent motion voting field output by S2 as the gating basis, for each candidate cross-frame connection, the following order is used for judgment and weighting: 1) The endpoint voxels are all located in the reachable space of potential dynamic targets and belong to the same risk corridor; 2) The angle between the displacement direction of the connection and the main direction of the risk corridor is calculated, and the angle does not exceed the direction tolerance recorded by the risk corridor; 3) Crossing detection is performed on the line segment connecting the centers of the endpoint voxels. If the line segment intersects with the static occlusion volume, the connection is rejected; 4) The visibility-consistent motion voting field values ​​in the direction of the connection are queried at the midpoint and two endpoints of the line segment, and the minimum value is used as the gating value of the connection. If it is lower than the gating threshold, it is rejected; 5) For the connection that passes the gating, its connection strength is adjusted according to the voting uncertainty. The greater the voting uncertainty, the lower the connection strength; the cross-frame connection after the above gating and weighting is used to aggregate the micro-motion evidence chain along the risk corridor in S4.

5. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The process of aggregating cross-frame occupancy changes and minute displacements along the risk corridor to form a consistent micro-motion evidence chain is as follows: In each risk corridor, a voxel located in the reachable space of a potential dynamic target and undergoing cross-frame occupancy changes is selected as the starting node, and the evidence chain is generated recursively along the cross-frame connecting edge according to the following rules: 1) Both voxels at both ends of the connecting edge are located within the risk corridor; 2) The consistency between the displacement direction of the connecting edge and the main direction of the risk corridor at that location satisfies the direction tolerance; 3) The connecting edge segment does not intersect with the static occlusion volume. 4) When a short invisible segment appears, the empty segment is only allowed to be inserted in the order of adjacent frames in the neighborhood of the occlusion entrance / exit, and the recursion can only continue after the pairwise occupancy change is confirmed on both sides of the line of sight visibility boundary; 5) When multiple feasible edges appear, the joint sorting of edge strength and occupancy consistency weight is given priority, and the second-best branch is reserved for parallel recursion; the formed evidence chain records the voxel index and edge set in chronological order for subsequent aggregation; The provision allowing short invisible segments at obstructed entrances and exits, but requiring confirmation through paired occupancy changes on both sides of the visibility boundary, specifically means: when an observation gap appears in the chain of evidence at an obstructed entrance or exit due to its location on the invisible side, the crossing confirmation is made only if the length of the short invisible segment does not exceed the upper limit, based on paired occupancy changes on both sides of the visibility boundary; after confirmation, the invisible segment is incorporated into the temporal continuity of the chain of evidence according to the correspondence of boundary normal alignment, and this rule cannot be used in areas without obstructed entrances or exits; The output of the three-dimensional verification candidate and evidence strength is as follows: the evidence chains of micro-motions with consistent directions within the same risk corridor are aggregated according to the spatial and temporal overlap relationship to obtain the three-dimensional verification candidate; the three-dimensional center is determined by the weighted average of the voxel centers on the chain, the orientation is determined by the weighted synthesis of the displacement directions of the connecting edges on the chain, and the evidence strength is determined by the weighted sum of the connecting edge strengths on the chain according to the occlusion consistency weight; the three-dimensional verification candidate and evidence strength are used as inputs of S5 for the initial estimation of the restricted state.

6. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The constraint estimation of candidate center, velocity and orientation is specifically as follows: taking the three-dimensional verification candidate output by S4 as input, the following steps are taken and constrained in the corresponding risk corridor: (1) Center constraint: taking the voxel center on the evidence chain that constitutes the candidate as a sample, the initial value of the center is obtained by weighting the product of the edge strength and the occlusion consistency weight; if the initial value is located in the static occlusion volume or the impassable area, it is projected back to the reachable space of the potential dynamic target along the normal of the nearest risk corridor center line, and it is ensured that it does not intersect with the static occlusion volume; (2) Velocity constraint: taking the projection of the cross-frame connection edge in the evidence chain in the risk corridor direction as the velocity sample, retaining the connection edge whose angle with the risk corridor direction does not exceed the direction tolerance. The velocity vector is synthesized by weighting the edge strength and occlusion consistency weight; the normal component of the velocity is truncated with the upper limit corresponding to the direction tolerance, and the final velocity is mainly projected in the direction of the risk corridor; (3) Orientation constraint: the unit direction of the velocity vector is used as the initial value of orientation, and fine adjustment is allowed within the ± direction tolerance range to adapt to the local curvature of the risk corridor; if the orientation causes the displacement line segment to intersect with the static occlusion volume as time progresses, it will retreat to the nearest direction that satisfies the non-intersection constraint; (4) Consistency gating: for the above combination of center, velocity and orientation, it is required that it is within the risk corridor and does not cross the static occlusion volume, and is consistent with the direction of the motion voting field with visibility consistency. Combinations below the gating threshold are eliminated; The use of occlusion consistency weight in the weighting process is as follows: In the weighted synthesis of center, velocity, and orientation, the edge strength of each cross-frame connection in the evidence chain constituting the 3D verification candidate is multiplied by the occlusion consistency weight of its spatial unit as the sample weight. Samples with higher weights contribute more to the estimation results. When the same candidate involves multiple risk corridor segments, only the orientation tolerance and occlusion consistency weight recorded in the risk corridor segment to which the candidate belongs are used for weighting to avoid cross-segment interference.

7. The method for identifying and tracking occluded dynamic targets based on deep learning according to claim 1, characterized in that... : The motion voting field based on visibility consistency extends along the risk corridor direction, and the extension terminates when it reaches the line-of-sight visibility boundary and there is a static occlusion volume in front. Specifically, it is as follows: Based on the initial estimate of state S5, the main direction and velocity magnitude are determined in the current risk corridor. An extension direction is selected along the main direction within the direction tolerance range. The step size is the time interval between adjacent frames multiplied by the velocity magnitude. The extension is carried out near the center line of the risk corridor by the step size. The direction consistency value of the motion voting field with visibility consistency is queried on the advancement line segment. It is required that the value is not lower than the gate threshold. During the advancement process, the line segment is continuously constrained to be completely located in the risk corridor and not intersect with the static occlusion volume. When the ray along the extension direction reaches the line-of-sight visibility boundary first and there is a static occlusion volume in front of it, the extension is terminated at the line-of-sight visibility boundary and marked as the extension endpoint. Otherwise, the extension continues until the next observation segment appears or the termination condition is met, provided that the above gate and constraint conditions are met. The process of reproducing the evidence chain when a new micro-motion evidence chain appears in a subsequent frame and is consistent with the original risk corridor is as follows: continuously monitor the micro-motion evidence chain output by S4 in the neighborhood of the extended endpoint and connect it according to the following consistency rules: (1) Spatial consistency: the distance between the starting voxel center of the evidence chain and the extended endpoint does not exceed the bandwidth of the risk corridor at that location, and both are located in the reachable space of the potential dynamic target; (2) Directional consistency: the angle between the displacement direction of the connecting edge of the evidence chain and the main direction of the risk corridor at that location does not exceed the directional tolerance, and is in the same direction as the extension direction; (3) Visibility consistency support: the minimum value of the line segment connecting the extended endpoint and the starting point of the evidence chain in the visibility consistency motion voting field is not lower than the gating threshold. (4) Obstruction geometry consistency: The above connecting line segments do not cross the static obstruction volume and conform to the lateral relationship of the line of sight visibility boundary; when all rules are met, the evidence link is added to the trajectory and the extension segment is replaced with the observation data of the evidence chain to achieve complete reproduction; when there are multiple candidate evidence chains, the one with the highest evidence strength is connected first, and the unselected candidate does not change the current trajectory.

Images