Target tracking method and device of vehicle, electronic equipment and storage medium

Abstract

A target tracking method and device of a vehicle, electronic equipment and a storage medium are disclosed, and relate to the technical field of vehicle environment perception. The method comprises: detecting the position and motion parameters of a target of the vehicle; in the case where first detection data of the target is obtained, fusing second detection data of the target on the basis of the first detection data to obtain fused target data; the different detection data is obtained by different sensor modules detecting the position and motion parameters of the target, and the priority of the first detection data is higher than that of the second detection data; determining the motion state of the target according to the fused target data, and tracking the target according to the motion state. The scheme discards the traditional approximate synchronization waiting mechanism, takes the arrival of the first detection data with higher priority as the only fusion trigger event, greatly reduces the fusion and tracking delay, and improves the reliability of target tracking.

Description

Technical Field

[0001] This application relates to the field of vehicle environmental perception, and more specifically, to a vehicle target tracking method, apparatus, electronic device, and storage medium. Background Technology

[0002] Currently, in the field of autonomous driving environmental perception, it is necessary to output information such as the three-dimensional position, size, orientation, speed, and category of surrounding targets in real time and accurately within the vehicle coordinate system. Performing target tracking based on fused data collected from multiple sensors is a key means to achieve high-precision perception. Specifically, existing multi-source data fusion schemes typically rely on an approximate synchronization mechanism using message filters, waiting for the timestamps of multiple detection results to align before target tracking. However, this approximate synchronization mechanism leads to increased waiting time and overall latency when sensor frequencies are inconsistent or communication jitter occurs. Furthermore, when any detection source drops a frame or has no output, fusion failure is likely to occur, affecting continuous target tracking and system robustness. Summary of the Invention

[0003] In view of this, embodiments of this application propose a vehicle target tracking method, apparatus, electronic device, and storage medium to improve the above-mentioned problems.

[0004] According to a first aspect of the embodiments of this application, a vehicle target tracking method is provided, the method comprising: detecting the position and motion parameters of a target in the vehicle; upon obtaining first detection data of the target, fusing second detection data of the target based on the first detection data to obtain fused target data; wherein the first detection data is obtained by a first set of sensors in the vehicle detecting the position and motion parameters of the target, the second detection data is obtained by a second set of sensors in the vehicle detecting the position and motion parameters of the target, and the first detection data has a higher priority than the second detection data; determining the motion state of the target based on the fused target data, and tracking the target based on the motion state.

[0005] In some embodiments, determining the motion state of the target based on the fused target data includes: calculating motion state determination information based on multiple sets of fused target data of the target over continuous time, wherein the motion state determination information includes at least one of the target's position change and velocity estimate; and determining the motion state of the target based on the motion state determination information.

[0006] In some embodiments, tracking the target based on the motion state includes: when the motion state indicates that the target is in a moving state, executing a full tracking strategy to track the target, wherein the full tracking strategy is constructed based on a nonlinear motion model and a Kalman filter algorithm based on unscented transform; and when the motion state indicates that the target is in a stationary state, executing a simplified tracking strategy to track the target, wherein the simplified tracking strategy is used to indicate a portion of the full tracking strategy.

[0007] In some embodiments, the second detection data includes measurement noise feature information of the target. The step of fusing the second detection data of the target based on the first detection data to obtain fused target data includes: matching the first detection data and the second detection data based on a preset multi-dimensional cost matrix to determine a set of targets to be tracked; the multi-dimensional cost matrix includes at least one of spatial distance, three-dimensional intersection-union ratio, category constraint, and sensor noise; the target set includes at least one candidate target; the detection data of the candidate target simultaneously belongs to both the first detection data and the second detection data; and the at least one candidate target includes the target. The motion parameters of each candidate target are merged based on the measurement noise feature information to obtain fused target data.

[0008] In some embodiments, matching the first detection data and the second detection data based on a preset multi-dimensional cost matrix includes: performing coordinate unification processing on the first detection data and the second detection data; and matching the coordinate-unified first detection data and the second detection data based on the multi-dimensional cost matrix.

[0009] In some embodiments, merging the motion parameters of each candidate target based on the measurement noise feature information to obtain fused target data includes: performing a weighted average of the motion parameters of each candidate target based on the measurement noise feature information of each candidate target as a weight to obtain fused target data.

[0010] In some embodiments, the first detection data includes the timestamp of the target, and the second detection data includes the timestamp of the target. The step of obtaining the first detection data and the second detection data includes: based on the timestamp of the first detection data, extracting the latest second detection data from a preset cache whose difference from the timestamp is less than or equal to a preset delay threshold.

[0011] According to a second aspect of the present application, a vehicle target tracking device is provided. The device includes: a target detection module for detecting the position and motion parameters of a target within the vehicle; a data fusion module for fusing second detection data of the target based on the first detection data to obtain fused target data, wherein the first detection data is obtained by a first set of sensors in the vehicle detecting the position and motion parameters of the target, and the second detection data is obtained by a second set of sensors in the vehicle detecting the position and motion parameters of the target, and the first detection data has a higher priority than the second detection data; and a tracking module for determining the motion state of the target based on the fused target data and tracking the target based on the motion state.

[0012] According to a third aspect of the present application, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, wherein when executed by the processor, the computer-readable instructions implement the vehicle target tracking method described above.

[0013] According to a fourth aspect of the embodiments of this application, a vehicle is provided, including: a body and an electronic device as described above; the electronic device is disposed inside the body.

[0014] According to a fifth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer-readable instructions thereon, which, when executed by a processor, implement the vehicle target tracking method described above.

[0015] According to a sixth aspect of the present application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium; the computer instructions are read and executed by a processor of a computer device to implement the vehicle target tracking method described above.

[0016] In the scheme of this application, the arrival of the first detection data with higher priority is used as the only fusion trigger event, which abandons the approximate synchronous waiting mechanism of the traditional message filter. The fusion process does not need to wait for the results of the second detection source to be collected, realizing low-latency multi-source data fusion and target tracking, and improving the reliability of target tracking.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the embodiments of this application. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0019] Figure 1 This is a schematic diagram of the structure of a target tracking system according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a priority-triggered target fusion module according to an embodiment of this application; Figure 3 This is a flowchart illustrating a priority-triggered target fusion module according to an embodiment of this application; Figure 4 This is a flowchart illustrating the CTRV-UKF fusion tracking module according to an embodiment of this application; Figure 5 This is a schematic flowchart illustrating a vehicle target tracking method according to an embodiment of this application; Figure 6 This is a schematic flowchart illustrating a vehicle target tracking method according to another embodiment of this application; Figure 7 This is a schematic flowchart illustrating a vehicle target tracking method according to another embodiment of this application; Figure 8 This is a schematic flowchart of a vehicle target tracking method according to another embodiment of this application; Figure 9 This is a schematic flowchart of a vehicle target tracking method according to another embodiment of this application; Figure 10 This is a schematic flowchart of a vehicle target tracking method according to another embodiment of this application; Figure 11 This is a block diagram of a vehicle target tracking device according to an embodiment of this application; Figure 12 This is a hardware structure diagram of an electronic device according to an embodiment of this application.

[0020] The accompanying drawings have illustrated specific embodiments of the present application. More detailed descriptions will follow. These drawings and descriptions are not intended to limit the scope of the present application's embodiments in any way, but rather to illustrate the concepts of the present application's embodiments to those skilled in the art through specific embodiments. Detailed Implementation

[0021] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0022] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0023] To facilitate understanding of the technical solution of this application, the overall architecture of the target tracking system involved in this application will first be described below. Figure 1 As shown, the system can include an onboard computing platform, LiDAR, camera, millimeter-wave radar, point cloud main detection module, point cloud clustering detection module, return on investment (ROI) constrained clustering detection module, radar target processing module, priority-triggered target fusion module, stationary target detection and processing module, CTRV-UKF fusion tracking module, tracked target output module, and downstream trajectory prediction and planning module. It integrates the priority-triggered low-latency fusion mechanism, CTRV model-unscented Kalman filter high-precision tracking algorithm, multi-sensor noise fine-grained weighted fusion, and dynamic and static target differentiated processing strategy into a unified design. It retains the multimodal perception architecture with LiDAR point cloud 3D detection as the main detection source. Through deep coupling of algorithms and architecture, it achieves low-latency, high-reliability, high-precision, and low-computing-power autonomous driving environment perception.

[0024] Alternatively, each module of the system can be deployed as an independent ROS / ROS2 node, running on an Ubuntu system on an in-vehicle computing platform. Nodes can interact with each other via topics, as shown in the following deployment: 1. Perception hardware nodes: LiDAR node ( / points_raw), camera node ( / image_raw), millimeter-wave radar node ( / radar_raw); 2. Multi-source detection nodes: point cloud main detection node ( / det / main_3d), point cloud clustering detection node ( / det / clus_3d), ROI constrained clustering detection node ( / det / roi_3d), radar target processing node ( / det / radar_3d); 3. Fusion Node: Priority-triggered target fusion node ( / fusion / multi_source_fuse); 4. Tracking nodes: Static target detection node ( / track / static_detect), CTRV-UKF fusion tracking node ( / track / ctrv_ukf_track); 5. Output node: Tracked target publication node ( / track / tracked_objects).

[0025] All nodes follow the common standards of the Autoware / APollo open-source autonomous driving framework in terms of message format, achieving seamless compatibility with downstream prediction and planning nodes.

[0026] Optionally, all core system parameters can be configurable through YAML configuration files, supporting dynamic adjustments at runtime. The core configuration items are divided into the following 5 categories, with example configurations as follows: 1. Fusion parameters: Buffer duration T = 0.3 "s", acquisition time threshold Δt' = 0.1 "s", Hungarian association cost weights α = 0.6, β = 0.3, γ = 0.1; 2. UKF tracking parameters: initial state vector value, covariance matrix P_0, process noise Q, Kalman gain threshold K_th=0.8; 3. Parameters for determining static and dynamic targets: Static velocity threshold V_static = 0.2 m / s, Position change threshold P_static = 0.1 m, Static update frequency f_static = 2 Hz; 4. Sensor calibration parameters: radar installation offset, camera intrinsic / extrinsic parameters, lidar-camera joint calibration parameters; 5. Computing power configuration parameters: number of CPU / GPU cores, memory usage threshold, and inference accuracy (FP32 / FP16 / INT8) for each module.

[0027] Optionally, the system's processing strategy can be dynamically adjusted based on the real-time computing load of the onboard computing platform: 1. High computing load (load > 80%): Disable the ROI constraint clustering detection module, retain only the main point cloud detection and point cloud clustering detection, and reduce the update frequency of stationary targets to 1Hz; 2. Medium computing load (50%≤load≤80%): Retain all detection sources, update stationary targets at 2Hz, and update dynamic targets at 10Hz; 3. Low computing load (load < 50%): Retain all detection sources, enable radar fusion, and update all targets at 10Hz.

[0028] The computing power scheduling is implemented by the vehicle-mounted computing power management node. By monitoring the CPU / GPU utilization, it sends dynamic configuration instructions to each module to ensure that the system can still run stably in computing power-constrained scenarios and meet real-time requirements.

[0029] For example, the 3D detection results output by the point cloud main detection module are set as high-priority main detection results. Machine learning 3D detection algorithms such as CenterPoint, TransFusion-L, and BEVFusion-L are used, or a combination of "2D instance segmentation + 3D shape estimation" is used to achieve the output of a main detection target list containing target category, 3D position, size, orientation, motion speed and timestamp, providing core basic data for fusion and tracking.

[0030] For example, the point cloud clustering detection module performs Euclidean clustering on the non-ground point cloud after ground removal to generate candidate 3D targets; the ROI-constrained clustering detection module projects the LiDAR point cloud onto the camera image plane, extracts the point cloud within the region based on the 2D detected ROI, and clusters it to generate enhanced 3D targets, compensating for the shortcomings of the main detection in small targets and occluded scenes; the radar target processing module performs coordinate system transformation, denoising, and quality assessment on the polar coordinate target information output by the millimeter-wave radar, outputting Cartesian coordinate radar candidate targets containing relative velocity and Doppler information, while also providing auxiliary basis for target motion state determination. Each secondary detection source runs independently and in parallel, and the output results are all marked with timestamps and sensor measurement noise, and are uploaded to the priority-triggered target fusion module in real time.

[0031] For example, combined Figure 2 The priority-triggered target fusion module may include a main trigger unit, a result caching unit, a collection unit, an association and merging unit, and a publishing unit. The association and merging unit may include a coordinate unification subunit, a Hungarian algorithm association subunit, and a noise inverse weighted fusion subunit. Combined with... Figure 3The priority-triggered target fusion module eliminates the near-synchronous waiting of message filters in traditional triggering sequences. It uses the arrival of the main detection result as the sole fusion trigger event and the point cloud main detection output sequence as the priority triggering sequence. After the main detection result is triggered, the main triggering unit immediately initiates a fusion process. Even if secondary detection sources are missing, there is no need to wait for all secondary detection source results to arrive, thus eliminating the latency caused by synchronous waiting. Each secondary detection source is configured with an independent result caching unit, using a circular buffer or timestamp index structure to store the detection output within the most recent time window. When fusion is triggered, the acquisition unit extracts the latest valid result from the cache of each secondary detection source that is closest to the current main detection timestamp and does not exceed a preset delay threshold. If a secondary detection source has no valid result, it is determined that the input is missing, and this fusion skips that input without affecting the overall fusion process. The coordinate unification subunit and the Hungarian algorithm association subunit first transform the 3D position, velocity, orientation, and other parameters of all primary and secondary detected targets to the vehicle's Cartesian coordinate system. Then, a multi-dimensional cost matrix is ​​constructed, incorporating spatial distance, 3D intersection-over-union (IoU), class constraints, and sensor noise. The Hungarian algorithm is then used to achieve high-precision one-to-one association between primary and secondary detected targets, solving the matching error problem of traditional simple association. For successfully associated target pairs, the noise inverse weighting fusion subunit uses an inverse noise weighting strategy to initially merge the motion parameters of the target pairs. The inverse value of the sensor measurement noise covariance is used to characterize the measurement accuracy; higher accuracy results in greater fusion weights. Unmatched secondary detected targets are included as new candidate targets in the fusion result. The final output unit provides a unified fused target set with standardized parameters, sensor noise labels, and association confidence, providing standardized input for subsequent tracking algorithms.

[0032] For example, combined Figure 4This application integrates the Constant Turn Rate and Velocity (CTRV) model and the Unscented Kalman Filter (UKF) as the core algorithms for target tracking into a dedicated CTRV-UKF fusion tracking module. For newly detected targets output by the fusion module, based on their standardized motion parameters (position, velocity, and orientation) and sensor noise labels, an independent UKF tracker instance and a unique tracking identity ID are assigned to each target. The UKF state vector and state covariance matrix are initialized, and the process noise covariance is configured as the noise variance of longitudinal acceleration and yaw angle acceleration according to the characteristics of the vehicle scene, realizing personalized initialization of the tracker. Based on the CTRV nonlinear motion model, the optimal tracking state of the previous frame is used as a basis, and the time step is determined according to the sampling frequency of the main detection. Prior prediction of the target state in the current frame is performed to obtain the prior state and prior covariance. There is no need to linearize the nonlinear model, which fundamentally avoids linearization error and is suitable for nonlinear motion scenarios such as vehicle turning and lane changing. The standardized fusion target set output by the fusion module is used as the measurement input for UKF. A measurement noise matrix is ​​constructed based on the measurement noise covariance of each sensor. The prior state is mapped to the measurement space through unscented transformation. The predicted measurement value, measurement covariance, and state-measurement cross-covariance are calculated, and then the Kalman gain is solved. The prior state is corrected using the measurement residuals to obtain the posterior optimal tracking state and posterior covariance of the current frame, achieving fine-grained fusion and updating of multi-sensor measurement values. For the geometric parameters of the target, such as length, width, height, and 3D height, a noise inverse weighted fusion strategy consistent with the motion parameters is adopted. The results of geometric parameter detection from the primary and secondary detection sources are continuously updated to ensure the accuracy and consistency of the geometric parameters.

[0033] For example, the stationary target detection and processing module is used as a pre-trigger unit for CTRV-UKF fusion tracking to achieve differentiated processing of stationary and moving targets. This reduces computational overhead while ensuring tracking accuracy. It fuses multi-dimensional information such as position changes across multiple consecutive frames, UKF estimated velocity, radar Doppler velocity, and target category priors to determine the target's motion state. If the position change is less than a preset threshold, the estimated velocity is close to zero, and there is no dynamic radar feedback, the target is determined to be stationary; otherwise, it is determined to be moving. Targets with unknown states are temporarily treated as moving targets. For targets determined to be moving, a complete CTRV-UKF prediction-update process is executed, retaining all dimensions of the state vector and updating all motion parameters such as position, velocity, orientation, and yaw rate in real time. Weighted fusion of full measurements from multiple sensors ensures high-precision tracking in nonlinear motion scenarios. For targets determined to be stationary, a lightweight simplified UKF tracking strategy is implemented, fixing the yaw rate and velocity, and simplifying the state vector to update only the position and orientation; at the same time, the update frequency is reduced, the shape update logic is simplified, and the tracker lifecycle threshold is extended, reducing unnecessary computation and resource consumption, and the computing power overhead is reduced by more than 60% compared to full tracking.

[0034] For example, the CTRV-UKF fusion tracking module has a built-in tracking lifecycle management subunit. It sets rules for tracker persistence, suspected disappearance, and deletion based on the UKF posterior covariance and target matching degree: when the posterior covariance exceeds a preset threshold, it is marked as a low-confidence tracking target; when no fusion-detected target is matched for multiple consecutive frames, it is marked as a suspected disappearance target; when the suspected disappearance state continues to exceed the threshold, the tracker is automatically deleted to avoid invalid calculations. Finally, the tracking target output module encapsulates the unique ID, 3D position, size, orientation, motion speed, motion state (dynamic / static), and tracking confidence (inferred from the posterior covariance) of all valid tracked targets into a standardized TrackedObject set, which is output to the downstream trajectory prediction and motion planning module. The interface is compatible with the general requirements of existing autonomous driving systems.

[0035] Figure 1 The system in [the document] can be used to implement the following Figure 5 For the vehicle target tracking method described, please refer to [link / reference]. Figure 5 , Figure 5 This application illustrates a vehicle target tracking method according to an embodiment of the present application. In a specific embodiment, this target tracking method can be applied to, for example... Figure 11 The vehicle target tracking device 300 and the electronic device 400 equipped with the vehicle target tracking device 300 are shown. Figure 12The specific process of the embodiments of this application will be described below. Of course, it is understood that this method can be executed by an electronic device with computing capabilities, such as an in-vehicle computing platform, a vehicle-side server, a cloud server, or other processors. The following will focus on... Figure 5 The process shown is described in detail. The target tracking method for the vehicle described above may specifically include the following steps S210-S230.

[0036] Step S210: Detect the target position and motion parameters of the vehicle.

[0037] For example, multiple sets of sensors deployed on a vehicle can be used to perceive and collect data on targets in the surrounding environment. It should be understood that these multiple sets of sensors may include lidar, cameras, and millimeter-wave radar. In other embodiments, they may also be ultrasonic radar, infrared sensors, or perception data provided by roadside units, as long as they can output detection data containing motion parameters such as target position, speed, and orientation. These sensors operate independently and in parallel, continuously outputting detection data streams with timestamps and their own measurement noise markers, providing a multi-source input basis for subsequent fusion and tracking.

[0038] It should be noted that the target mentioned here refers to various objects in the current vehicle's environment, which may include, but are not limited to, other moving vehicles, pedestrians, roadside guardrails, lampposts, other stopped vehicles, etc.

[0039] Step S220: After obtaining the first detection data of the target, the second detection data of the target is fused based on the first detection data to obtain fused target data.

[0040] The first detection data (corresponding to the main detection result output by the aforementioned main detection source) is obtained by the first set of sensors in the vehicle (such as the aforementioned main detection source) detecting the position and motion parameters of the target. The second detection data (corresponding to the secondary detection result output by the aforementioned secondary detection source) is obtained by the second set of sensors in the vehicle (such as the aforementioned secondary detection source) detecting the position and motion parameters of the target. The first detection data has a higher priority than the second detection data.

[0041] It should be noted that the number of secondary detection sources is unlimited and can be set according to actual needs. For example, in one alternative example, the number of secondary detection sources is one, which can be a camera. In another alternative example, the number of secondary detection sources is multiple, which can include LiDAR, cameras, and millimeter-wave radar. Each secondary detection source operates independently and in parallel, performing parallel detection of the target and outputting detection results in parallel.

[0042] For example, the core of this step lies in the "priority-triggered" fusion logic. Traditional multi-source fusion schemes typically rely on message filters to wait for the results of multiple detection sources to be approximately synchronized in timestamps before triggering fusion. This mechanism can lead to increased waiting time and overall latency when sensor frequencies are inconsistent or communication jitter occurs. Furthermore, fusion failure is likely to occur when any detection source drops a frame or has no output. This embodiment, however, abandons this near-synchronous waiting mechanism, using the arrival of the highest-priority first detection data as the sole fusion trigger event. Once the first detection data arrives, a fusion process is immediately initiated, extracting and merging usable second detection data without waiting for the results of the second detection sources to be complete. This mechanism eliminates the latency caused by synchronous waiting at its source, significantly reducing the overall processing latency of fusion and tracking, and improving system real-time performance. Simultaneously, when a second detection data source is occasionally missing due to occlusion, dropped frames, or algorithm failure, that input can be skipped directly, and fusion can be completed solely based on the first detection data or other available second detection data. This avoids the fusion failure or tracking interruption problems caused by missing inputs in traditional synchronous mechanisms, ensuring the continuity of target output and the robustness of the system. It should be understood that the distinction between first and second detection data is not absolutely fixed to a certain type of sensor. In different computing platforms or sensing architectures, the output of any detection source that is set as high priority and high reliability can be used as the first detection data, while the output of other auxiliary or supplementary detection sources can be used as the second detection data.

[0043] Step S230: Determine the motion state of the target based on the fused target data, and track the target based on the motion state.

[0044] For example, after obtaining the fused target data, this step does not apply an indiscriminate tracking strategy to all targets. Instead, it first determines the target's motion state based on the fused target data. Here, motion state includes both moving and stationary states. In complex autonomous driving scenarios, a large number of dynamic targets (such as moving vehicles and pedestrians) and stationary targets (such as roadside guardrails, lampposts, and stopped vehicles) exist simultaneously in the environment. Applying a tracking algorithm of equal complexity to all targets would result in an extremely high computational burden, making it difficult to meet the real-time requirements of the onboard computing platform. Therefore, this embodiment introduces motion state determination to provide a basis for subsequent differentiated tracking strategies. It should be noted that when the target's motion state is moving, continuous tracking of the target is required.

[0045] In this embodiment, the arrival of the first detection data with higher priority is used as the sole fusion trigger event. This eliminates the near-synchronous waiting mechanism of traditional message filters. The fusion process does not need to wait for the results of the second detection source to be collected, thus realizing low-latency multi-source data fusion and target tracking, and improving the reliability of target tracking.

[0046] Furthermore, after step S210, the embodiments of this application may also include the step of obtaining first detection data and second detection data. This step may include: extracting the latest second detection data from a preset cache based on the timestamp of the first detection data, the difference between the timestamp and the second detection data being less than or equal to a preset delay threshold.

[0047] For example, since the second detection data comes from different sensor groups, their output frequencies, communication delays, and computing power often differ. If a traditional approximate synchronous waiting mechanism is used, all detection sources must be timed to coincide before fusion can be triggered. This can lead to severe latency buildup or even fusion failure when sensors drop frames or communication jitter occurs. This application's embodiments abandon this synchronous waiting and configure an independent preset buffer for each channel of second detection data. When the highest-priority first detection data arrives, the latest second detection data, whose difference from the timestamp carried by the first detection data is less than or equal to a preset delay threshold, is immediately extracted from each preset buffer. For example, if the preset delay threshold is set to 0.1 seconds, only the latest valid results with a timestamp difference within 0.1 seconds are extracted. For second detection data sources exceeding this threshold or without valid results, they are determined to be occasional missing data and the input is skipped directly, continuing to perform subsequent fusion based on existing available data. It should be understood that the specific value of the preset latency threshold can be dynamically configured based on the real-time computing load of the onboard computing platform and the nominal frequency of the sensor. For example, the threshold can be appropriately relaxed under low computing load to accommodate greater communication jitter, while the threshold can be tightened under high computing load to further reduce fusion latency. This extraction logic based on timestamps and latency thresholds is mechanistically compatible with the occasional absence of secondary detection sources, ensuring that the fusion process is uninterrupted and the target output is continuously available, thus achieving a balance between low latency and high robustness.

[0048] Optionally, the specific method for obtaining the fused target data in step S220 is not limited and can be set according to actual application requirements. For example, in an alternative example, the second detection data may include the target's measurement noise characteristic information, combined with... Figure 6 Step S220 may include steps S221 and S222.

[0049] Step S221: Match the first detection data with the second detection data based on the preset multi-dimensional cost matrix to determine the target set to be tracked. The multi-dimensional cost matrix includes at least one of spatial distance, three-dimensional intersection-union ratio, category constraint and sensor noise. The target set includes at least one candidate target. The detection data of the candidate target belongs to both the first detection data and the second detection data. At least one candidate target includes the target.

[0050] It should be noted that traditional simple association methods (such as nearest neighbor association) rely only on a single spatial distance dimension, which is prone to mismatches in scenarios with dense targets, occlusion, or overlapping movements. This application's embodiments introduce a multi-dimensional cost matrix, comprehensively considering multiple dimensions such as spatial distance, 3D intersection-union ratio (IU), category constraints, and sensor noise. Specifically, the spatial distance term measures the Euclidean distance between the center points of the detection boxes, the IU term measures the overlap of the detection box volumes, the category constraint term imposes a penalty cost on out-of-class targets, and the sensor noise term incorporates measurement uncertainty into the association cost consideration. Through this multi-dimensional joint constraint, the globally optimal set of one-to-one association matching pairs can be solved, i.e., the set of targets to be tracked, where candidate targets possess attribute information from both the first and second detection data. It should be understood that the weight coefficients of the multi-dimensional cost matrix can be adaptively adjusted according to the perception characteristics of different sensors. For example, in sparse point cloud scenarios, the weight of category constraints can be increased, and in high-speed motion scenarios, the cost weight based on speed consistency can be increased, thereby ensuring the robustness and accuracy of association matching in complex scenarios.

[0051] Optionally, the specific matching method in step S221 is not limited and can be set according to actual application requirements. For example, in an alternative example, combining... Figure 7 Step S221 may include steps S2211 and S2212.

[0052] Step S2211: Perform coordinate unification processing on the first detection data and the second detection data.

[0053] Step S2212: Match the first detection data and the second detection data after coordinate unification based on the multi-dimensional cost matrix.

[0054] It should be noted that the raw detection data from different groups of sensors are typically output based on their respective local coordinate systems (e.g., LiDAR is based on the radar coordinate system, cameras on the image pixel coordinate system, and millimeter-wave radar on the polar coordinate system). To perform target association and merging under a unified spatial dimension, the three-dimensional position, velocity, orientation, and other motion parameters of all first and second detection data are uniformly transformed to the vehicle's Cartesian coordinate system. It should be understood that although the embodiments of this application preferably unify the coordinates to the vehicle coordinate system, in other embodiments, they can also be unified to the global navigation coordinate system or the local coordinate system of the roadside unit, as long as the multi-source data is measurable under the same reference system. Coordinate unification eliminates spatial reference differences between multi-source data, laying the foundation for subsequent high-precision association and matching. After coordinate unification, the first and second detection data are associated with high precision one-to-one using the Hungarian algorithm, solving the matching error problem of traditional simple association.

[0055] Step S222: Based on the measurement noise feature information, the motion parameters of each candidate target are merged to obtain fused target data.

[0056] For example, after determining the set of targets to be tracked, the motion parameters of each candidate target are refined and merged. The measurement noise feature information (e.g., measurement noise variance) contained in the second detection data characterizes the sensor's measurement accuracy of the parameters.

[0057] As an alternative implementation, the motion parameters of each candidate target can be weighted and averaged based on the measurement noise characteristics of each candidate target to obtain fused target data.

[0058] For example, in this embodiment, the reciprocal of the measured noise feature information (i.e., the reciprocal of the noise variance) is used as the weight to perform a weighted average of the same motion parameter for each candidate target. It should be noted that sensors with smaller measured noise variance have higher measurement accuracy and should be given greater weight during fusion; conversely, sensors with larger noise variance have higher measurement uncertainty and should be given less weight. Through this collaborative weighting strategy of giving greater weight to higher accuracy sensors, the fusion result can maximize the absorption of the observation advantages of high-precision sensors while suppressing the bias introduced by low-precision sensors, achieving refined matching and collaborative enhancement of multi-sensor measurement characteristics.

[0059] Optionally, the specific method for determining the motion state in step S230 is not limited and can be set according to actual application requirements. For example, in an alternative example, combined with Figure 8 Step S230 may include steps S231 and S232.

[0060] Step S231: Based on multiple sets of fused target data over continuous time, calculate motion state determination information, wherein the motion state determination information includes at least one of the target's position change and velocity estimate.

[0061] It should be noted that after obtaining multiple sets of fused target data over continuous time, the judgment is not made based solely on a single set of data, but rather by comprehensively considering the target's motion characteristics in the time dimension. The change in position is characterized by calculating the difference in the target's three-dimensional coordinates across multiple consecutive sets to represent the absolute magnitude of its spatial displacement; the velocity estimate is directly derived from the velocity parameters in the fused target data, or derived from the ratio of the change in position to the time step. It should be understood that the motion state determination information is not limited to the change in position and the velocity estimate. In other embodiments, multi-dimensional information such as Doppler radial velocity feedback provided by radar sensors and prior probability of the target category (e.g., the known pedestrian category has a higher dynamic probability) can also be introduced to improve the robustness of the judgment in occluded or sparse point cloud scenarios.

[0062] Step S232: Determine the motion state of the target based on the motion state determination information.

[0063] It should be noted that a hard determination of the target's motion state is achieved by comparing the calculated motion state determination information with a preset threshold. To illustrate this scheme more clearly, an example threshold is given below. It should be understood that this value is only an example and not a limitation: when the velocity estimate is less than 0.2 m / s and the change in the fused target position for three consecutive sets is less than 0.1 m, the target is determined to be in a stationary state; otherwise, it is determined to be in a moving state. For newly detected targets that have not yet accumulated continuous data, they are temporarily designated as moving by default. The stationary state is switched to after multiple consecutive sets of data meet the stationary threshold. This conservative strategy avoids unnecessary full-tracking computational overhead caused by misjudging stationary targets as dynamic targets due to initial observation noise.

[0064] Optionally, the specific method of target tracking in step S230 is not limited and can be set according to actual application requirements. For example, in an alternative example, combining... Figure 9 Step S230 may also include steps S233 and S234.

[0065] Step S233: When the motion state indicates that the target is in a moving state, execute the full tracking strategy to track the target. The full tracking strategy is constructed based on a nonlinear motion model and a Kalman filter algorithm based on unscented transform.

[0066] It should be noted that, in combination Figure 10 For dynamic targets determined to be in a moving state, the system enters the dynamic target branch and executes a full tracking strategy. In autonomous driving scenarios, vehicle movements such as turning and lane changing are inherently nonlinear. If traditional classical Kalman filtering or extended Kalman filtering is used, the former is only applicable to linear systems, while the latter, although it approximates nonlinear models through Jacobian matrix linearization, truncates higher-order terms during the linearization process, introducing a non-negligible truncation error, leading to a sharp drop in tracking accuracy in strongly nonlinear scenarios. This embodiment employs a Kalman filtering algorithm based on a nonlinear motion model and unscented transformation (e.g., constant turning rate and speed model CTRV combined with unscented Kalman filtering UKF). Its core mechanism is that unscented transformation captures the statistical characteristics of random variables through a set of defined sampling points (Sigma points), directly substituting these sampling points into the nonlinear motion model for mapping, without requiring any linearization approximation or differentiation operations on the model itself, thus fundamentally avoiding linearization truncation errors and achieving high-precision state estimation in nonlinear scenarios.

[0067] For example, the specific prediction and update steps of the full tracking strategy are as follows: First, during the tracker initialization phase, the state vector and covariance matrix are initialized based on the fused target data. The state vector contains all motion parameters of the target, such as position, velocity, orientation, and angular velocity.

[0068] The prediction step then begins: based on the posterior optimal state of the previous frame, a set of augmented sampling points is generated; next, these sampling points are mapped to the current time step using a nonlinear motion model to obtain the mapped prior sampling points; finally, the prior sampling points are weighted to obtain the prior state estimate and prior covariance estimate for the current time step. This process completely preserves the physical reality of the nonlinear motion.

[0069] Finally, the update step begins: The fused target measurement value and its measurement noise covariance are used as input. Prior sampling points are mapped to the measurement space, and the predicted measurement value, measurement covariance, and state-measurement cross-covariance are calculated. Then, the Kalman gain is solved, and the prior state is corrected using the measurement residuals to obtain the posterior optimal tracking state and posterior covariance. Through this closed-loop prediction and update process, the full-scale tracking strategy can accurately capture the nonlinear trajectory changes of dynamic targets.

[0070] Step S234: When the motion state indicates that the target is stationary, a simplified tracking strategy is executed to track the target. The simplified tracking strategy is used to indicate a part of the full tracking strategy.

[0071] It should be noted that, in combination Figure 10 For stationary targets determined to be stationary, the system enters the stationary target branch and executes a simplified tracking strategy. This "simplified strategy" is not an arbitrary reduction, but rather a targeted computational optimization based on the physical characteristics of stationary targets. The velocity and angular velocity of a stationary target are theoretically always zero. If a full tracking strategy were still applied, not only would the velocity and angular velocity dimensions in the state vector continuously introduce meaningless prediction noise, leading to unnecessary expansion of the covariance matrix, but the high-frequency update operations would also consume a large amount of computing resources from the onboard computing platform.

[0072] To illustrate this solution more clearly, a simplified implementation of the tracking strategy is given below. It should be understood that this approach is merely an example and not a limitation: In terms of simplifying the state vector dimension, the simplified tracking strategy reduces the 5-dimensional state vector in the full tracking strategy to 3-dimensional, retaining only the position and orientation parameters, and directly fixing the velocity and angular velocity parameters to 0, thereby significantly compressing the dimension and computational cost of the state covariance matrix and the process noise matrix.

[0073] In terms of simplifying the update frequency, the full tracking strategy usually updates at a high frequency (e.g., 10Hz) based on the frequency of the main detection source, while the simplified tracking strategy reduces the update frequency to 2Hz, that is, only one filtering update is performed every 5 frames, and the remaining frames are only used for position observation buffering without triggering matrix operations.

[0074] In terms of simplifying the state covariance matrix, the state covariance matrix is ​​simplified to 3-dimensional, and only the noise of position and orientation is updated.

[0075] In terms of simplifying lifecycle management, considering the characteristic that stationary targets usually exist for a long time, the simplified tracking strategy extends the tracker's keep-alive threshold to 100 frames, reducing the overhead of trackers being frequently created and destroyed due to brief occlusion.

[0076] By simplifying the state vector dimension, update frequency, state covariance matrix, and lifecycle as described above, the simplified tracking strategy significantly reduces the computational overhead of a single stationary target while ensuring the accuracy of monitoring the position and orientation of the stationary target. This application's embodiments implement a partially customized strategy based on physical characteristics, achieving an optimal balance between accuracy and computational power. This allows for the simultaneous fulfillment of high-precision tracking requirements for dynamic targets and low-overhead monitoring requirements for stationary targets with limited computational resources.

[0077] To more clearly illustrate the technical advantages of the above-mentioned solutions in this application, the embodiments of this application will be described in detail with reference to two typical application scenarios.

[0078] Scenario 1 is an application scenario of multi-source detection and CTRV-UKF fusion tracking without millimeter radar fusion. The main detection source is 3D detection of LiDAR point cloud, and the secondary detection sources are point cloud clustering detection and ROI constraint clustering detection. The link is: LiDAR / camera → multi-source detection module → priority triggered target fusion module → stationary target detection and processing module → CTRV-UKF fusion tracking module → tracking target output module.

[0079] For example, the CenterPoint / / BEVfusion-L / TransFusion-L machine learning 3D detection algorithm can be used to perform feature encoding and target detection on the LiDAR point cloud, outputting a list of main detected targets. The feature parameters of a single target are represented by the following formula 1:

[0080] Formula 1 in, This indicates the three-dimensional position in the vehicle coordinate system. Indicates the length, width, and height of the target. Indicates the target yaw angle. Indicates the absolute velocity of the target. Indicates the target category (e.g., vehicle, pedestrian, non-motorized vehicle). Represents a timestamp. This represents the variance of the main detection measurement noise. This indicates the unique identifier of the main detection target.

[0081] The point cloud clustering detection module performs Euclidean clustering on the non-ground point cloud after ground removal, generating point cloud clusters and estimating 3D bounding boxes using geometric rules, outputting a list of clustered detection targets. The ROI-constrained clustering detection module projects the point cloud onto the camera image plane, extracts and clusters point clouds based on 2D ROI detection, and outputs a list of ROI-enhanced detection targets. The target parameter formats of the two modules are completely consistent with the main detection module, differing only in the measurement noise variance and clustering detection noise variance. ROI Enhancement Detection Noise Variance .

[0082] The priority-triggered target fusion module uses the arrival of the main detection result as the trigger signal. After buffer acquisition, coordinate unification, Hungarian correlation, and preliminary fusion with inverse noise weighting, it outputs a unified fused target list. ,for , Configure a circular buffer unit, cache duration When the main detection result Upon arrival, retrieve the cached data with the current timestamp. Time difference The latest secondary test results are as follows: , ,and , .

[0083] In this embodiment, all detection sources have been initially converted to the vehicle coordinate system. Only the position and velocity of secondary detection targets are finely calibrated. The calibration formula is as follows: Formula 2: Formula 2 in, , To calibrate the offset for the position, The velocity calibration coefficients are used to obtain the standardized secondary target parameters after calibration. , .

[0084] For static targets (such as roadside fixed calibration targets, guardrails, and streetlights), multiple sets of target position data from primary and secondary detection are collected, and the optimal translation calibration parameters are solved using the least squares method. Formula 3 minimizes the mean square error (MSE) between the calibrated secondary detection position and the primary detection position: Formula 3 in, For the number of calibration samples, For secondary detection The location of each sample The location of the sample corresponding to the main detection; For dynamic targets (such as test vehicles traveling at a constant speed), multiple sets of speed data from primary and secondary detection are collected, and the optimal speed calibration coefficient is solved using the least squares method. Formula 4 minimizes the mean square error between the calibrated secondary speed and the primary detection speed. Formula 4 In the embodiments of this application, the following are taken This is the optimal coefficient obtained after dynamic calibration, used to correct the speed underestimation / overestimation bias of secondary detection.

[0085] A multi-dimensional cost matrix is ​​constructed to associate the primary detection target with the secondary detection target. A single cost element is defined by Formula 5: Formula 5 Among them, spatial distance item Represents the normalized value of the three-dimensional Euclidean distance. , To achieve the maximum detection distance, normalized to 3D IoU item This represents the intersection-union ratio (IUR) of the 3D bounding boxes of the primary and secondary targets. , For the intersection volume, For the union volume, the category penalty term Similar targets outlier targets , , , Weighting coefficient: In this embodiment, we take... , , ,satisfy .

[0086] The minimum weight matching of the cost matrix is ​​obtained by solving the Hungarian algorithm, thus obtaining the set of matching pairs. Unmatched main detection target set Unmatched secondary detection target set The matching rules are as follows: If a match is found, it is considered a valid match; otherwise, it is considered a non-match.

[0087] For the set of matching pairs The target in the equation is obtained by applying inverse noise weighting to the motion parameters using Formula 6. Initial fusion will be carried out: Formula 6 The fusion formula is the same as the above formula, only the corresponding parameters and noise variance are replaced; geometric parameters ( ) Targets will not be merged for now and will be left to be updated by the tracking module; unmatched targets will be retained directly in the fusion result.

[0088] Final output fusion target list A single objective parameter is represented by the following formula 7: Formula 7 in, The variance of the noise after fusion. , .

[0089] A two-dimensional judgment method is used to determine the dynamic and static states of the fusion target. The judgment rules are as follows: Speed ​​determination: Fusion speed ; Position change determination: Position change of three consecutive sets of fused targets ; If both of the above conditions are met, it is a stationary target, and is marked as such. Otherwise, it is a dynamic target and is marked as... The new detection target is temporarily marked as (Dynamic), switching to [a new mode] after 3 consecutive sets of data meet the conditions. In this embodiment of the application, two dynamic targets (vehicles) and one stationary target (roadside guardrail) were detected, and were marked as follows: , , .

[0090] The CTRV-UKF fusion tracking module assigns an independent UKF tracker to each fusion target. Dynamic targets undergo full CTRV-UKF fusion tracking, while stationary targets undergo simplified CTRV-UKF fusion tracking. The core algorithm flow is: tracker initialization → UKF prediction step (time update) → UKF update step (measurement update) → geometric parameter fusion. The following uses a dynamic vehicle target as an example to illustrate the fusion target parameters. Initialize the UKF state vector to the initial values. State covariance matrix Process noise covariance matrix .

[0091] The state vector (CTRV model, 5-dimensional) is represented by the following formula 8: Formula 8 In this embodiment, the following is taken ( (For initial straight-line driving).

[0092] The state covariance matrix (a diagonal matrix with the same dimension as the state vector) is represented by the following formula 9: Formula 9 In this embodiment, we take .

[0093] Process noise covariance matrix (2D, corresponding to longitudinal acceleration) Yaw angle acceleration This is expressed by the following formula 10: Formula 10 In this embodiment, we take It matches the motion noise of the CTRV model.

[0094] Based on the CTRV nonlinear motion model, sigma points are generated through unscented transformation (UT), and the prior state is obtained after mapping. with prior covariance The core steps and formulas are as follows: Generate augmented sigma points: augmented state Augmented covariance Sampling parameters ( To augment the state dimension), then ; Sigma point set There are 15 points in total, and the generation formula is as follows: Formula 11: Formula 11 CTRV model mapping to sigma points: For the motion parameters of each sigma point, the CTRV discrete-time motion model is used to map them to time k. In this embodiment... (For straight-line driving), the mapping formula simplifies to the following formula 12: Formula 12 like Then the complete CTRV mapping formula is used as shown in Formula 13 below: Formula 13 Weighted calculation of prior state and covariance: The formula for sigma point weights is shown in Equation 14 below: Formula 14 Substitution , ,have to , ( ),satisfy Prior state: (Taking the first 5 dimensions of motion parameters); Prior covariance: .

[0095] In this embodiment, the prior state at time k=1 is calculated as follows:

[0096] Prior covariance:

[0097] fusion target parameters As measured value Combined with the measurement noise covariance matrix Correcting the prior state yields the posterior optimal state. With posterior covariance The core formula is as follows: Measured values ​​and measurement noise: Measured values (4-dimensional, taking the core motion parameters of the fusion target); Measurement noise covariance matrix The variance is obtained from the noise variance after fusion.

[0098] The sigma point is mapped to the measurement space as shown in Equation 15 below: Formula 15 in, For identity mapping, motion parameters are directly extracted; The predicted measured values ​​and the measurement covariance are shown in Equations 16 and 17 below: Formula 16 Formula 17 The state-measure cross covariance is shown in Equation 18 below: Formula 18 Kalman gain and posterior state update are shown in Equations 19, 20, and 21 below: Formula 19 Formula 20 Formula 21 In this embodiment, the posterior optimal state at time k=1 is calculated as follows:

[0099] Posterior covariance Comparison A 30% reduction indicates an improvement in tracking confidence.

[0100] Length, width and height of the target The geometric parameters of multi-source detection are obtained by inverse weighted fusion of noise using the following formula 22: Formula 22 The fusion formula is consistent with the above formula, and in this embodiment, we obtain... , , .

[0101] For stationary targets (roadside guardrails), perform lightweight processing: The state vector is simplified to 3 dimensions: ,fixed , ; The update frequency has been reduced to 2Hz (once every 5 frames). The state covariance matrix is ​​simplified to 3D, and only the noise of position and orientation is updated; Extend the lifecycle keep-alive threshold to 100 frames to reduce tracker creation / destruction overhead.

[0102] The posterior optimal motion parameters of dynamic targets, combined with fused geometric parameters, and the simplified tracking parameters of stationary targets are encapsulated into a standardized TrackedObject set. The parameters of a single tracked target are shown in the following formula 23: Formula 23 in, To track confidence level ( (For matrix traces), in this embodiment, the dynamic target confidence level Confidence of stationary targets The output is sent to the downstream trajectory prediction module.

[0103] Scenario 2 is a multi-source detection and CTRV-UKF fusion tracking application scenario that includes radar fusion. Based on Scenario 1, millimeter-wave radar is added as an optional secondary detection source. A new radar target processing module and velocity consistency association constraints are added to enhance the velocity estimation of dynamic targets and the ability to perceive adverse scenes. The core improvements are radar target coordinate transformation, velocity consistency constraints of multi-source fusion, and updated radar velocity assistance in CTRV-UKF. The rest is the same as Scenario 1.

[0104] Millimeter-wave radar outputs a list of polar coordinate targets. The single target parameter is ,in The radar detection range is in meters (m). It is the radar azimuth angle (rad, the angle between the radar X-axis and the azimuth angle). The radar radial velocity (m / s) is used; the radar target processing module converts it into target parameters in the vehicle's Cartesian coordinate system. The core conversion formula is shown in Formula 24 below: Formula 24 in, Install azimuth calibration coefficients on the radar. , To calibrate the offset of the radar installation location, To fuse the target's yaw angle, radar measurement noise variance After conversion, the radar target parameter format is consistent with other detection sources, and a radar candidate target list is output. .

[0105] In the Hungarian algorithm association cost matrix of Scenario 1, a speed consistency term is added. The optimization cost element formula is shown in Formula 25 below: Formula 25 Among them, speed consistency item , To achieve maximum detection speed, normalized to The weighting coefficients are adjusted to , , , If the weight sum is 1, the weight of the velocity term of the radar target is increased to [value missing]. This enhances the advantages of speed measurement.

[0106] In the UKF update step, the radar speed will be... Added to the measurement vector as an independent measurement value This expands the measurement vector to 5 dimensions, as shown in Equation 26 below: Formula 26 Simultaneously measure the noise covariance matrix Expanded to 5 dimensions, the noise variance corresponding to radar velocity is taken as... With all other parameters unchanged, after UKF update, the target velocity estimation error is reduced by 45% compared to scenario one, significantly improving the accuracy and stability of dynamic target velocity tracking.

[0107] The above embodiments describe in detail the vehicle target tracking method provided in this application. In other embodiments, this application also provides a vehicle target tracking device. Figure 11 This is a block diagram of a vehicle target tracking device according to an embodiment of this application, such as... Figure 11 As shown, the target tracking device 300 for the vehicle includes: a target detection module 310, a data fusion module 320, and a tracking module 330.

[0108] The target detection module 310 is used to detect the position and motion parameters of the target on the vehicle; The data fusion module 320 is used to fuse the second detection data of the target based on the first detection data of the target to obtain fused target data. The first detection data is obtained by the first set of sensors in the vehicle detecting the position and motion parameters of the target, and the second detection data is obtained by the second set of sensors in the vehicle detecting the position and motion parameters of the target. The first detection data has a higher priority than the second detection data. The tracking module 330 is used to determine the motion state of the target based on the fused target data, and to track the target based on the motion state.

[0109] In other embodiments, this solution also provides an electronic device, such as... Figure 12 As shown, the electronic device 400 also includes a processor 410 and one or more memories 420, the one or more memories 420 being used to store program instructions executed by the processor 410, the processor 410 implementing the above-described vehicle target tracking method when executing the program instructions.

[0110] Furthermore, the processor 410 may include one or more processing cores. The processor 410 runs or executes instructions, programs, code sets, or instruction sets stored in the memory 420, and retrieves data stored in the memory 420. Optionally, the processor 410 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 410 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor and may be implemented using a separate communication chip.

[0111] In other embodiments, the present solution also provides a vehicle that may include a body and electronic devices as described above, the electronic devices being disposed inside the body.

[0112] In other embodiments, this solution also provides a computer-readable storage medium, which may be included in the cloud server described in the above embodiments; or it may exist independently and not assembled into the cloud server. The aforementioned computer-readable storage medium carries computer-readable instructions, which, when executed by a processor, implement the methods in any of the above embodiments.

[0113] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, but not limited to: electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0114] In an exemplary embodiment, this application also provides a computer program product, which, when executed by a processor, is used to implement the above-described vehicle target tracking method.

[0115] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0116] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0117] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A vehicle target tracking method, characterized in that, include: The target position and motion parameters of the vehicle are detected; If the first detection data of the target is obtained, the second detection data of the target is fused based on the first detection data to obtain fused target data; Wherein, the first detection data is obtained by the first set of sensors in the vehicle detecting the position and motion parameters of the target, the second detection data is obtained by the second set of sensors in the vehicle detecting the position and motion parameters of the target, and the first detection data has a higher priority than the second detection data; The motion state of the target is determined based on the fused target data, and the target is tracked based on the motion state.

2. The method according to claim 1, characterized in that, Determining the motion state of the target based on the fused target data includes: Based on multiple sets of fused target data over continuous time, motion state determination information is calculated, wherein the motion state determination information includes at least one of the target's position change and velocity estimate. The motion state of the target is determined based on the motion state determination information.

3. The method according to claim 2, characterized in that, The step of tracking the target based on the motion state includes: When the motion state indicates that the target is in a moving state, a full tracking strategy is executed to track the target, wherein the full tracking strategy is constructed based on a nonlinear motion model and a Kalman filter algorithm based on unscented transform; When the motion state indicates that the target is stationary, a simplified tracking strategy is executed to track the target, wherein the simplified tracking strategy is used to indicate a portion of the full tracking strategy.

4. The method according to claim 1, characterized in that, The second detection data includes the measurement noise characteristic information of the target. The step of fusing the second detection data of the target based on the first detection data to obtain fused target data includes: The first detection data and the second detection data are matched based on a preset multi-dimensional cost matrix to determine the target set to be tracked. The multi-dimensional cost matrix includes at least one of spatial distance, three-dimensional intersection-union ratio, class constraint and sensor noise. The target set includes at least one candidate target. The detection data of the candidate target belongs to both the first detection data and the second detection data. The at least one candidate target includes the target. Based on the measured noise characteristic information, the motion parameters of each candidate target are merged to obtain fused target data.

5. The method according to claim 4, characterized in that, The matching of the first detection data and the second detection data based on a preset multi-dimensional cost matrix includes: The first detection data and the second detection data are processed to unify their coordinates; The first and second detection data after coordinate unification are matched based on a multi-dimensional cost matrix.

6. The method according to claim 4, characterized in that, The process of merging the motion parameters of each candidate target based on the measured noise feature information to obtain fused target data includes: Based on the measurement noise feature information of each candidate target as weights, the motion parameters of each candidate target are weighted and averaged to obtain fused target data.

7. The method according to claim 1, characterized in that, The first detection data includes the timestamp of the target, and the second detection data includes the timestamp of the target. The steps of obtaining the first detection data and the second detection data include: Based on the timestamp of the first detection data, the latest second detection data with a difference from the timestamp less than or equal to a preset delay threshold is extracted from the preset cache.

8. A vehicle target tracking device, characterized in that, The device includes: The target detection module is used to detect the position and motion parameters of targets on the vehicle. The data fusion module is used to fuse the second detection data of the target based on the first detection data of the target to obtain fused target data, provided that the first detection data of the target has been obtained. Wherein, the first detection data is obtained by the first set of sensors in the vehicle detecting the position and motion parameters of the target, the second detection data is obtained by the second set of sensors in the vehicle detecting the position and motion parameters of the target, and the first detection data has a higher priority than the second detection data; The tracking module is used to determine the motion state of the target based on the fused target data, and to track the target based on the motion state.

9. An electronic device, characterized in that, The electronic device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 7.

10. A vehicle, characterized in that, include: The vehicle body and the electronic device of claim 9; the electronic device is disposed inside the vehicle body.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 7.

12. A computer program product, characterized in that, The computer program product includes computer instructions, which are stored in a computer-readable storage medium. The computer instructions are read and executed by the processor of the computer device to implement the method as described in any one of claims 1 to 7.

Images