Construction method of expressway train dynamic online map based on sparse perception

Abstract

This invention relates to a method for constructing dynamic online maps of highway trains based on sparse perception, belonging to the field of autonomous driving and vehicle-road cooperative technology. To address the shortcomings of existing technologies in online map construction for multi-vehicle platooning scenarios on highways, this invention employs a sparse modeling approach to extract only necessary map elements, reducing computational and communication overhead, while supporting synchronous updates of dynamic road elements. It is particularly suitable for multi-vehicle collaborative online map construction in highway platooning scenarios. With the development of V2X and vehicle-road cooperative perception technologies, multiple autonomous vehicles can share environmental information, improving the overall perception range and robustness. This invention, combined with the highway platooning scenario, proposes a sparse perception-driven dynamic online map construction method by fusing visual sensor data and motion states from multiple vehicles, enabling real-time map construction collaboratively by vehicle platoons.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and vehicle-road cooperation technology, specifically to a method for constructing a dynamic online map of highway trains based on sparse perception. Background Technology

[0002] Traditional autonomous driving systems primarily rely on high-definition maps (HD maps) or dense bird's-eye view (BEV) features for environmental mapping. HD maps require regular updates by specialized data collection vehicles, resulting in high costs and low freshness, making it difficult to reflect dynamic road changes (such as temporary construction or obstacles) in a timely manner. While online mapping methods based on dense BEVs reduce reliance on prior maps, they incur significant computational overhead. Converting multi-view images to BEVs requires complex depth estimation, leading to decreased real-time performance and failing to meet the demands of high-speed scenarios. Furthermore, existing solutions are mostly designed for single-vehicle systems, failing to effectively utilize the advantages of multi-vehicle collaboration in platooning scenarios: each vehicle repeatedly collects data on the same road segments, causing redundant computation and communication burdens, and failing to address core issues such as multi-vehicle perspective complementarity and spatiotemporal alignment.

[0003] In recent years, some studies have attempted to optimize mapping efficiency and accuracy. Patent CN202411222915.7, a vectorized map construction method based on graph structure modeling, enhances map element representation through subgraph node attention, but remains limited to single-vehicle perception and does not address multi-vehicle data fusion. Patent CN202411210817.1, an online mapping method combining navigation maps, introduces lane-level structured information, but relies on prior map matching, failing in areas without pre-stored maps or when road conditions change abruptly, and does not support dynamic target tracking. These methods have significant shortcomings in multi-vehicle platooning scenarios on highways: high computational load, delayed dynamic element updates, and poor robustness of multi-source data fusion, hindering the real-time performance and reliability of platoon collaborative perception.

[0004] To address the aforementioned bottlenecks, this invention proposes a sparse perception-driven multi-vehicle collaborative online mapping mechanism: it reduces computational and communication overhead by compressing map representation through polyline anchor points; it achieves continuous tracking of dynamic targets and efficient deduplication of static elements by combining temporal caching and cross-vehicle coordinate alignment; and it significantly improves mapping efficiency and freshness in high-speed scenarios without the need for prior map support. Summary of the Invention

[0005] In view of the above problems, this invention provides a method for constructing a dynamic online map of highway trains based on sparse perception. To address the shortcomings of existing online map construction methods in multi-vehicle platoon scenarios on highways, this invention proposes a dynamic online mapping method based on multi-vehicle data fusion. The algorithm employs sparse modeling to extract only necessary map elements, reducing computational and communication overhead, while supporting the synchronous updating of dynamic road elements (such as vehicle trajectories).

[0006] This invention specifically addresses multi-vehicle collaborative online map construction technology for highway platooning scenarios. With the development of vehicle-to-everything (V2X) and vehicle-road cooperative perception technologies, multiple autonomous vehicles can share environmental information, improving overall perception range and robustness. Combining the highway platooning scenario, this invention proposes a sparse perception-driven dynamic online map construction method by fusing visual sensor data and motion states from multiple vehicles, enabling real-time map construction collaboratively by the vehicle platoon.

[0007] This invention provides a method for constructing a dynamic online map of highway trains based on sparse sensing, comprising: Step S1: Acquire multi-view images of each vehicle in the highway train through onboard multi-view cameras, and input them into a pre-trained image encoder with shared weights to generate multi-scale image feature maps of each vehicle. Step S2: Input the multi-scale image feature map I of each vehicle into the Transformer decoder, perform dynamic target query and static map query respectively, and decode the sparse representation perception results of each vehicle, including static map elements and dynamic targets. Step S3: Construct a temporal memory queue to cache the historical sensing results of the most recent H frames; The target points in the historical frames of each vehicle are transformed to the coordinate system of the current frame through motion compensation. Then, the dynamic target query and static map query of the current frame are used as the query and the historical features after motion compensation are used as the key / value to perform multi-head cross-attention fusion, and the enhanced dynamic query and static query of each vehicle are output. The system assigns continuous IDs to dynamic targets using an ID matching header and outputs a list of targets with historical trajectories. Simultaneously, it performs temporal completion and confidence updates on static polylines. Finally, it outputs a list of dynamic targets with continuous IDs and historical trajectories for each vehicle, a list of completed static polylines for each vehicle, and the current frame self-pose of each vehicle, for subsequent multi-vehicle communication and fusion. Step S4: Encode and compress the dynamic target list with continuous ID and historical trajectory for each vehicle, and the completed static polyline list for each vehicle, and broadcast them to other vehicles in the platoon via the vehicle V2V communication link. Based on its relative pose transformation matrix with the sending vehicle, the receiving vehicle transforms the coordinates of the enhanced sparse sensing result of the current frame to its own local coordinate system, thus obtaining a set of static polylines and a set of dynamic targets in a unified coordinate system for each vehicle. Step S5: For the static polyline set and the dynamic target set under the unified coordinate system of each vehicle, perform static element fusion, dynamic target fusion and confidence fusion respectively to generate the global static map element set and the global dynamic target status list after fusion of each vehicle. Step S6: Combine the global static map feature set after fusion of each vehicle and the global dynamic target status list after fusion of each vehicle, and combine the vehicle's pose in the current frame to generate a structured dynamic online map. This map includes a static polyline structure with road edges, lane lines, and guardrails, and a dynamic trajectory sequence with the historical and predicted trajectories of surrounding vehicles. The map is then presented in a structured format for use by the vehicle controller.

[0008] Optionally, step S1 includes the following specific steps: Step S11: The multi-view camera system consists of 6 to 8 cameras per vehicle, distributed in front, behind, left, and right of each vehicle to form an omnidirectional view. Step S12: Input the multi-view images of each vehicle into the ResNet50+FPN image encoding network, and output multi-scale feature maps. , where s represents the scale and n represents the camera number; Step S13: Perform fusion processing on the multi-scale feature maps to form a unified feature space, enhance the expressive ability of key structures in the image, and obtain multi-scale image feature map I.

[0009] Optionally, the dynamic target query outputs the 3D bounding box parameters of the dynamic target; The sparse static map includes static road polyline anchor points and their corresponding feature vectors; wherein, the polyline anchor points are composed of a series of ordered discrete point coordinates, the number of which is much smaller than the number of pixels of the corresponding map elements, thereby achieving sparse representation.

[0010] Optionally, step 3 may include the following specific steps: Step 3.1: Each autonomous vehicle maintains a first-in-first-out temporal memory queue in its local memory to cache the historical perception results of the most recent H frames. Each frame of data in the queue includes: The timestamp of this frame is synchronized by GPS or PTP protocol with an accuracy in the microsecond range; The vehicle's pose at this frame moment is provided by RTK-GPS / IMU, including position and orientation; The dynamic target list obtained by decoding this frame includes 3D bounding box parameters, temporary IDs, confidence scores, and feature vectors. The static map target list obtained by decoding this frame, namely the polyline anchor points, contains a set of ordinal points, type labels, confidence scores, and feature vectors.

[0011] After processing the current frame, the data of that frame is packaged and pushed to the end of the queue; if the queue is full and the length has reached H, the oldest frame at the head of the queue, i.e. the TH frame, is discarded to ensure that the queue always stores the data of the most recent H frames. Step 3.2: Transform the spatial points of all dynamic targets and static map targets in each historical frame from the vehicle coordinate system of the historical frame to the vehicle coordinate system of the current frame to achieve motion compensation and obtain the features of each historical frame in the current vehicle coordinate system. Specifically, this includes: the pose of each historical frame in the current vehicle coordinate system, the transformed list of dynamic targets, the transformed list of static map targets, and the feature vector associated with each target point. The dynamic target query of the current frame is processed by multi-head cross-attention with the historical dynamic target features and historical static map features of each historical frame in the vehicle coordinate system, and the output is the current frame dynamic query carrying historical dynamic target information and the current frame static query carrying historical static map information. The current frame dynamic query carrying historical dynamic target information obtained in step 3.3 is sent into the ID matching header, and a list of dynamic targets with continuous IDs and historical trajectory associations is output. The current frame static query, carrying historical static map information, performs a polyline completion operation to obtain a static polyline list that has undergone time-series completion and confidence updates.

[0012] Optionally, step S4 includes the following specific steps: Step 4.1 Each autonomous vehicle performs lightweight encoding of its sparse perception results in the current frame and broadcasts them to other vehicles in the platoon via the V2V network. Step 4.2 The receiving vehicle continuously monitors the communication channel and receives data packets from other vehicles; For each received valid data packet, after decoding and timestamp verification, the receiver obtains the self-position pose of the sending vehicle and its own current self-position pose, and then performs coordinate transformation and alignment. The receiver obtains sparse map elements and dynamic target data from all other vehicles that have been unified into its own local coordinate system. Step 4.3 Synchronize the timestamp information of multi-vehicle data to ensure consistent processing in the time domain, and output a static polyline set and a dynamic target set in a unified coordinate system.

[0013] Optionally, step 6 includes the following specific steps: Step 6.1: Convert the deduplicated global static map feature set into a structured static map layer. Each static map feature is represented in the form of a polyline, specifically including: polyline point set, feature type label, confidence score, and additional attributes. Step 6.2: Convert the fused global dynamic target state list into a dynamic target layer. Each dynamic target contains: a globally unique ID, current frame state, historical trajectory, predicted trajectory, target type, and confidence level.

[0014] Step 6.3: Merge the static map layer and the dynamic target layer to form a complete dynamic online map data structure.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: (1) In view of the problem that existing high-precision maps are heavily dependent and slow to update, this invention adopts a sparse perception structure to build online maps in real time based on perception results, thereby getting rid of the dependence on prior maps and realizing on-demand map generation; (2) In view of the problem of high computational cost and high data volume of BEV feature mapping, this invention uses anchor point polylines and feature vectors to represent map elements, reducing unnecessary dense representation and communication burden, and improving perception and mapping efficiency; (3) In view of the problem of overlapping vision and duplicate data in multiple workshops, this invention proposes a multi-vehicle collaborative fusion and NMS deduplication mechanism, which combines coordinate transformation and confidence weighting to retain unique high-quality map elements and improve perception accuracy. (4) In view of the problem that it is difficult to continuously map the target when the target is dynamically changing, the present invention improves the continuity of target tracking by using a time-series cache queue and instance association mechanism to form a continuous target trajectory in the dynamic map; (5) In view of the difficulty of time and space alignment in multi-vehicle data fusion, this invention introduces a multi-vehicle anchor point coordinate projection mechanism and a timestamp synchronization mechanism based on relative pose to ensure map structure consistency and communication robustness. Attached Figure Description

[0016] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the flowchart of the sparse sensing module and the time-series cache queue in an embodiment of the present invention. Detailed Implementation To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0018] A specific embodiment of the present invention, such as Figure 1-2 A method for constructing a dynamic online map of highway trains based on sparse perception is disclosed. The specific implementation steps are as follows: Step S1: Acquire multi-view images of each vehicle in the highway train through onboard multi-view cameras, and input them into a pre-trained image encoder with shared weights to generate multi-scale image feature maps. It is understandable that high-speed trains are convoyed with high-speed vehicles; Optionally, step S1 includes the following specific steps: Step S11: The multi-view camera system consists of 6 to 8 cameras per vehicle, distributed in front, behind, left, and right of each vehicle to form an omnidirectional view. Step S12: Input the multi-view images of each vehicle into the ResNet50+FPN image encoding network, and output multi-scale feature maps. , where s represents the scale and n represents the camera number; Step S13: Perform fusion processing on the multi-scale feature maps to form a unified feature space, enhance the expressive ability of key structures in the image, and obtain multi-scale image feature map I.

[0019] Optionally, the key structures are lane lines, pedestrians, and edge obstacles.

[0020] Step S2: Input the multi-scale image feature map I of each vehicle into the Transformer decoder, perform dynamic target query and static map query respectively, and decode the sparse representation perception results of each vehicle, including static map elements and dynamic targets. Optionally, the dynamic target query outputs the 3D bounding box parameters of the dynamic target; The sparse static map includes static road polyline anchor points and their corresponding feature vectors; wherein, the polyline anchor points are composed of a series of ordered discrete point coordinates, the number of which is much smaller than the number of pixels of the corresponding map elements, thereby achieving sparse representation.

[0021] Optionally, the dynamic targets are vehicles and pedestrians and their corresponding positions, sizes, angles, and speeds; The sparse static map targets lane lines and guardrails; Optionally, the number of anchor points for the polyline can be 5 to 7.

[0022] Step S3: Construct a temporal memory queue to cache the historical sensing results of the most recent H frames; The target points in the historical frames of each vehicle are transformed to the coordinate system of the current frame through motion compensation. Then, the dynamic target query and static map query of the current frame are used as the query and the historical features after motion compensation are used as the key / value to perform multi-head cross-attention fusion, and the enhanced dynamic query and static query of each vehicle are output. The system assigns continuous IDs to dynamic targets using an ID matching header and outputs a list of targets with historical trajectories. Simultaneously, it performs temporal completion and confidence updates on static polylines. Finally, it outputs a list of dynamic targets with continuous IDs and historical trajectories for each vehicle, a list of completed static polylines for each vehicle, and the current frame self-pose of each vehicle, for subsequent multi-vehicle communication and fusion. Optionally, step 3 may include the following specific steps: Step 3.1: Each autonomous vehicle maintains a first-in-first-out temporal memory queue in its local memory to cache the historical perception results of the most recent H frames. Each frame of data in the queue includes: The timestamp of this frame is synchronized by GPS or PTP protocol with an accuracy in the microsecond range; The vehicle's pose at this frame moment is provided by RTK-GPS / IMU, including position and orientation; The dynamic target list obtained by decoding this frame includes 3D bounding box parameters, temporary IDs, confidence scores, and feature vectors. The static map target list obtained by decoding this frame, namely the polyline anchor points, contains a set of ordinal points, type labels, confidence scores, and feature vectors.

[0023] After processing the current frame, the data of that frame is packaged and pushed to the end of the queue; if the queue is full and the length has reached H, the oldest frame at the head of the queue, i.e. the TH frame, is discarded to ensure that the queue always stores the data of the most recent H frames. Step 3.2: Transform the spatial points of all dynamic targets and static map targets in each historical frame from the vehicle coordinate system of the historical frame to the vehicle coordinate system of the current frame to achieve motion compensation and obtain the features of each historical frame in the current vehicle coordinate system. Specifically, this includes: the pose of each historical frame in the current vehicle coordinate system, the transformed list of dynamic targets, the transformed list of static map targets, and the feature vector associated with each target point. The dynamic target query of the current frame is processed by multi-head cross-attention with the historical dynamic target features and historical static map features of each historical frame in the vehicle coordinate system, and the output is the current frame dynamic query carrying historical dynamic target information and the current frame static query carrying historical static map information. The current frame dynamic query carrying historical dynamic target information obtained in step 3.3 is sent into the ID matching header, and a list of dynamic targets with continuous IDs and historical trajectory associations is output. The current frame static query carrying historical static map information performs a polyline completion operation to obtain a static polyline list that has been completed in time series and updated with confidence. Optionally, each target includes 3D bounding box parameters, velocity, confidence level, and optionally a sequence of trajectory points from several past frames; Each polyline contains an ordered set of points, a type label, and a confidence level; Optionally, the specific steps include: Extract the features of each detected target in the current frame and the information of each target in historical frames; Calculate the cost matrix between the target in the current frame and the targets in historical frames. The Hungarian algorithm is used to find the optimal match and obtain the correspondence between the target in the current frame and the target in the historical frames.

[0024] For a successfully matched target, assign the ID from the historical frame to the target in the current frame; for a newly appearing target, assign a new global temporary ID; for a disappeared target, end its trajectory recording.

[0025] Optionally, the specific steps for performing the polyline completion operation include: merging the set of polyline points in the historical frame that belong to the same physical feature as the polyline in the current frame; Since historical frames may observe parts that are missing in the current frame due to occlusion, the corresponding polyline points in the historical frames can be added to the polyline of the current frame. The number of points in the completed polyline may exceed K, and it needs to be reduced back to K points through uniform sampling to maintain the consistency of the data structure; At the same time, the confidence level of the same line is dynamically adjusted based on the number of times it appears in historical frames: the more times it appears, the higher the confidence level.

[0026] Step S4: Encode and compress the dynamic target list with continuous ID and historical trajectory for each vehicle, and the completed static polyline list for each vehicle, and broadcast them to other vehicles in the platoon via the vehicle V2V communication link. Based on its relative pose transformation matrix with the sending vehicle, the receiving vehicle transforms the coordinates of the enhanced sparse sensing result of the current frame to its own local coordinate system, thus obtaining a set of static polylines and a set of dynamic targets in a unified coordinate system for each vehicle. Optionally, step S4 includes the following specific steps: Step 4.1 Each autonomous vehicle performs lightweight encoding of its sparse perception results in the current frame and broadcasts them to other vehicles in the platoon via the V2V network. This invention transmits only sparse map features and compact feature vectors, avoiding the transmission of original images or dense point cloud data, thereby significantly reducing communication bandwidth usage and transmission latency.

[0027] Step 4.2 The receiving vehicle continuously monitors the communication channel and receives data packets from other vehicles; For each received valid data packet, after decoding and timestamp verification, the receiver obtains the self-position pose of the sending vehicle and its own current self-position pose, and then performs coordinate transformation and alignment. The receiver obtains sparse map elements and dynamic target data from all other vehicles that have been unified into its own local coordinate system. Position and pose information includes latitude and longitude, altitude, heading angle, pitch angle, and roll angle; Step 4.3 Synchronize the timestamp information of multi-vehicle data to ensure consistent processing in the time domain, and output a static polyline set and a dynamic target set in a unified coordinate system.

[0028] Step S5: For the static polyline set and the dynamic target set under the unified coordinate system of each vehicle, perform static element fusion, dynamic target fusion and confidence fusion respectively to generate the global static map element set and the global dynamic target status list after fusion of each vehicle. Optionally, aligned map features from multiple vehicles can be aggregated in the local coordinate system of the receiving vehicle. For static map features, a non-maximum suppression (NMS) algorithm suitable for polyline structures is adopted. Matching is performed based on the spatial distance and shape similarity between polylines, retaining the polyline anchor point with the highest confidence and eliminating redundancy. For dynamic targets, fusion is performed through target ID matching or location association.

[0029] Optionally, the specific steps of the static element fusion include: Spatial pairing of polyline representations of static elements such as lane lines from different vehicles; An algorithm similar to Non-Maximum Suppression (NMS) is adopted. When multiple lines have similar shapes, the one with the highest confidence is retained to achieve deduplication and optimization.

[0030] Optionally, the specific steps of the dynamic target fusion include: Based on ID and location, different observations of the same dynamic target are fused to obtain a more accurate state estimate.

[0031] Confidence fusion assigns weights to the perception results of different vehicles (e.g., vehicles closer to the target have higher weights), thereby improving the reliability of the fusion results.

[0032] Optionally, step S5 is as follows: Step S5.1: A set of static map feature polylines from all vehicles, after coordinate alignment; Based on the spatial distance and shape similarity between polylines, polylines from different vehicles that describe the same map element are paired. For each pair of paired polylines, a non-maximum suppression algorithm is used to retain the polyline with the highest confidence and suppress the other redundant polylines, resulting in a deduplicated, high-quality set of global static map features. Step S5.2: A set of dynamic target trajectories from all vehicles, each with a unique ID and location information; matching the same dynamic target perceived by different vehicles based on the target's ID, location, and motion state; for the same target that is successfully matched, using Kalman filtering or weighted average algorithm to fuse its position, speed, and other information from multiple perspectives to obtain a more accurate and stable global dynamic target state list. Step S5.3: All perceived data and their original confidence levels in the above fusion process; Assign appropriate weights to the perception results of different vehicles; For example, vehicles closer to the target and those with higher sensor configurations have higher weights in their perception results. The weights are then used to adjust the state and confidence of the fused static elements and dynamic targets, resulting in a global environment representation that integrates multi-vehicle information, is optimized for deduplication, and has a reliable confidence assessment.

[0033] Step S6: Combine the global static map feature set after fusion of each vehicle and the global dynamic target status list after fusion of each vehicle, and combine the vehicle's pose in the current frame to generate a structured dynamic online map. This map includes a static polyline structure with road edges, lane lines, and guardrails, and a dynamic trajectory sequence with the historical and predicted trajectories of surrounding vehicles. The map is then presented in a structured format for use by the vehicle controller.

[0034] Optionally, step 6 includes the following specific steps: Step 6.1: Convert the deduplicated global static map feature set into a structured static map layer. Each static map feature is represented in the form of a polyline, specifically including: polyline point set, feature type label, confidence score, and additional attributes. Polyline point set: ordered three-dimensional spatial point coordinates (x, y, z), with a fixed number of points K or a variable length; Element type labels: such as lane lines, curbs, guardrails, and divider boundaries; Confidence score: The confidence score after fusion, which indicates the reliability of the element; Additional attributes: such as lane line color, width, and line type.

[0035] Step 6.2: Convert the merged global dynamic target state list into a dynamic target layer, where each dynamic target contains: Globally unique ID, current frame state, historical trajectory, predicted trajectory, target type, and confidence level.

[0036] Globally unique ID: Maintains consistency within the queue; Current frame state: 3D bounding box, velocity vector, acceleration; Historical trajectory: A sequence of trajectory points from the past N frames provided by a temporal memory queue, where each trajectory point contains a timestamp and location coordinates; Predicted trajectory: The short-term future trajectory output based on historical trajectories and motion models; Target types: such as cars, trucks, pedestrians, and unknown obstacles; Confidence level: The confidence level after fusion.

[0037] Step 6.3: Merge the static map layer and the dynamic target layer to form a complete dynamic online map data structure. This map has the following characteristics: Real-time performance: Outputs the map at the current moment every frame (10Hz), without relying on offline high-precision maps; Sparsity: It contains only the necessary map elements (road structure represented by polylines + discrete dynamic targets), resulting in high storage and transmission efficiency; Dynamics: The trajectory of a dynamic target updates over time, reflecting the real-time movement of road participants; Vehicle center: The origin of the coordinate system is the center of the rear axle of the current vehicle. The X-axis is forward, the Y-axis is left, and the Z-axis is upward, which is convenient for downstream modules to use directly.

[0038] Optionally, the generated global map is provided to the planning and control modules of this vehicle or other vehicles, and can be uploaded to the cloud for application and updating. This includes: in the cloud, matching and merging static elements (such as newly added temporary construction fences or changed lane lines) with high confidence in the online map and those differing from existing high-precision maps with the existing map to achieve dynamic incremental updates of the high-precision map. In this process, the main data controlled are the geometric shape (polyline vertices) and semantic attributes of static map elements, while dynamic target trajectories are used for traffic flow analysis and are generally not included in the persistent content of the high-precision map.

[0039] The standardized structure includes a polyline map structure and a dynamic target trajectory sequence.

[0040] The map is continuously output at each time frame and used by the vehicle or the convoy master vehicle for path planning and collaborative control.

[0041] The polyline map structure consists of road edges, guardrails, and lane lines; The dynamic target trajectory sequence is the motion trajectory of the target vehicle in the current frame and the previous N frames.

[0042] The dynamic online map outputs map data for use by the vehicle controller in route planning or fleet coordination control modules.

[0043] Map data can be uploaded synchronously to upper-level platforms (such as cloud map update services) to help generate local high-precision maps or navigation layers.

[0044] Through the above methods, this invention achieves multi-workshop collaborative online mapping in high-speed scenarios while ensuring sparse efficiency, thus meeting the needs of autonomous driving fleets.

[0045] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a dynamic online map of highway trains based on sparse sensing, characterized in that, include: Step S1: Acquire multi-view images of each vehicle in the highway train through onboard multi-view cameras, and input them into a pre-trained image encoder with shared weights to generate multi-scale image feature maps of each vehicle. Step S2: Input the multi-scale image feature map I of each vehicle into the Transformer decoder, perform dynamic target query and static map query respectively, and decode the sparse representation perception results of each vehicle, including static map elements and dynamic targets. Step S3: Construct a temporal memory queue to cache the historical sensing results of the most recent H frames; The target points in the historical frames of each vehicle are transformed to the coordinate system of the current frame through motion compensation. Then, the dynamic target query and static map query of the current frame are used as the query and the historical features after motion compensation are used as the key / value to perform multi-head cross-attention fusion, and the enhanced dynamic query and static query of each vehicle are output. The system assigns continuous IDs to dynamic targets using an ID matching header and outputs a list of targets with historical trajectories. Simultaneously, it performs temporal completion and confidence updates on static polylines. Finally, it outputs a list of dynamic targets with continuous IDs and historical trajectories for each vehicle, a list of completed static polylines for each vehicle, and the current frame self-pose of each vehicle, for subsequent multi-vehicle communication and fusion. Step S4: Encode and compress the dynamic target list with continuous ID and historical trajectory for each vehicle, and the completed static polyline list for each vehicle, and broadcast them to other vehicles in the platoon via the vehicle V2V communication link. Based on its relative pose transformation matrix with the sending vehicle, the receiving vehicle transforms the coordinates of the enhanced sparse sensing result of the current frame to its own local coordinate system, thus obtaining a set of static polylines and a set of dynamic targets in a unified coordinate system for each vehicle. Step S5: For the static polyline set and the dynamic target set under the unified coordinate system of each vehicle, perform static element fusion, dynamic target fusion and confidence fusion respectively to generate the global static map element set and the global dynamic target status list after fusion of each vehicle. Step S6: Combine the global static map feature set after fusion of each vehicle and the global dynamic target status list after fusion of each vehicle, and combine the vehicle's pose in the current frame to generate a structured dynamic online map. This map includes a static polyline structure with road edges, lane lines, and guardrails, and a dynamic trajectory sequence with the historical and predicted trajectories of surrounding vehicles. The map is then presented in a structured format for use by the vehicle controller.

2. The method for constructing a dynamic online map of highway trains based on sparse sensing according to claim 1, characterized in that, The specific steps of step S1 include: Step S11: The multi-view camera system consists of 6 to 8 cameras per vehicle, distributed in front, behind, left, and right of each vehicle to form an omnidirectional view. Step S12: Input the multi-view images of each vehicle into the ResNet50+FPN image encoding network, and output multi-scale feature maps. , where s represents the scale and n represents the camera number; Step S13: Perform fusion processing on the multi-scale feature maps to form a unified feature space, enhance the expressive ability of key structures in the image, and obtain multi-scale image feature map I.

3. The method for constructing a dynamic online map of highway trains based on sparse sensing according to claim 1, characterized in that, The dynamic target query outputs the 3D bounding box parameters of the dynamic target; The sparse static map includes static road polyline anchor points and their corresponding feature vectors; wherein, the polyline anchor points are composed of a series of ordered discrete point coordinates, the number of which is much smaller than the number of pixels of the corresponding map elements, thereby achieving sparse representation.

4. The method for constructing a dynamic online map of highway trains based on sparse sensing according to claim 1, characterized in that, Step 3 includes the following specific steps: Step 3.1: Each autonomous vehicle maintains a first-in-first-out temporal memory queue in its local memory to cache the historical perception results of the most recent H frames. Each frame of data in the queue includes: The timestamp of this frame is synchronized by GPS or PTP protocol with an accuracy in the microsecond range; The vehicle's pose at this frame moment is provided by RTK-GPS / IMU, including position and orientation; The dynamic target list obtained by decoding this frame includes 3D bounding box parameters, temporary IDs, confidence scores, and feature vectors. The static map target list obtained by decoding this frame, namely the polyline anchor points, contains a set of ordinal points, type labels, confidence scores, and feature vectors. After processing the current frame, the data of that frame is packaged and pushed to the end of the queue; if the queue is full and the length has reached H, the oldest frame at the head of the queue, i.e. the TH frame, is discarded to ensure that the queue always stores the data of the most recent H frames. Step 3.2: Transform the spatial points of all dynamic targets and static map targets in each historical frame from the vehicle coordinate system of the historical frame to the vehicle coordinate system of the current frame to achieve motion compensation and obtain the features of each historical frame in the current vehicle coordinate system. Specifically, this includes: the pose of each historical frame in the current vehicle coordinate system, the transformed list of dynamic targets, the transformed list of static map targets, and the feature vector associated with each target point. The dynamic target query of the current frame is processed by multi-head cross-attention with the historical dynamic target features and historical static map features of each historical frame in the vehicle coordinate system, and the output is the current frame dynamic query carrying historical dynamic target information and the current frame static query carrying historical static map information. The current frame dynamic query carrying historical dynamic target information obtained in step 3.3 is sent into the ID matching header, and a list of dynamic targets with consecutive IDs and historical trajectory associations is output. The current frame static query, carrying historical static map information, performs a polyline completion operation to obtain a static polyline list that has undergone time-series completion and confidence updates.

5. The method for constructing a dynamic online map of highway trains based on sparse sensing according to claim 1, characterized in that, The specific steps of step S4 include: Step 4.1 Each autonomous vehicle performs lightweight encoding of its sparse perception results in the current frame and broadcasts them to other vehicles in the platoon via the V2V network. Step 4.2 The receiving vehicle continuously monitors the communication channel and receives data packets from other vehicles; For each received valid data packet, after decoding and timestamp verification, the receiver obtains the self-position pose of the sending vehicle and its own current self-position pose, and then performs coordinate transformation and alignment. The receiver obtains sparse map elements and dynamic target data from all other vehicles that have been unified into its own local coordinate system. Step 4.3 Synchronize the timestamp information of multi-vehicle data to ensure consistent processing in the time domain, and output a static polyline set and a dynamic target set in a unified coordinate system.

6. The method for constructing a dynamic online map of highway trains based on sparse sensing according to claim 1, characterized in that, Step 6 includes the following specific steps: Step 6.1: Convert the deduplicated global static map feature set into a structured static map layer. Each static map feature is represented in the form of a polyline, specifically including: polyline point set, feature type label, confidence score, and additional attributes. Step 6.2: Convert the fused global dynamic target state list into a dynamic target layer. Each dynamic target contains: a globally unique ID, current frame state, historical trajectory, predicted trajectory, target type, and confidence level. Step 6.3: Merge the static map layer and the dynamic target layer to form a complete dynamic online map data structure.

Images