Laser point cloud mapping and matching system and method based on cross-country scene

Abstract

The invention relates to a laser point cloud mapping and matching method based on a cross-country scene, and the method comprises the steps: 1, collecting a cross-country environment point cloud, carrying out the filtering of the point cloud, removing a dynamic object point cloud, and forming real-time point cloud data; step 2, performing pose prediction estimation on the vehicle sensor data through Kalman filtering, and correcting the predicted pose to obtain final pose estimation; 3, adding real-time point cloud data with semantic tags to the local increment dense voxel semantic map through final pose estimation to form a local increment dense semantic map; 4, performing point cloud registration on the real-time point cloud data and the local increment dense semantic map through a semantic point cloud observation model, and obtaining a point cloud key frame according to point cloud registration and final pose estimation; and step 5, estimating the accurate pose of the vehicle according to the point cloud key frame and the pose information corresponding to the point cloud key frame, constructing a consistency map, realizing global consistency optimization of the map, and completing construction of a global semantic map.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to a laser point cloud mapping and matching system and method based on off-road scenarios. Background Art

[0002] Currently, SLAM (Simultaneous Localization and Mapping) technology has made significant progress in regular environments such as urban roads and campuses. However, in unstructured off-road scenarios, due to factors such as uneven roads, obstructions from vegetation, and blurred boundaries of drivable areas, traditional point cloud mapping and matching algorithms suffer from reduced accuracy and insufficient robustness, making it difficult to meet the high-reliability mapping and positioning requirements of off-road unmanned vehicles. Notable characteristics of off-road scenarios include large terrain undulations, blurred drivable areas, and highly dynamic environments. Existing point cloud mapping systems often use structures such as voxel grids and OctoMaps. Under off-road conditions, these systems face challenges such as difficulty mapping sparse point clouds, missing key structural features, and matching drift. Therefore, a highly robust point cloud mapping and matching method suitable for off-road scenarios is urgently needed. Summary of the Invention

[0003] In view of the above problems, the present invention provides a laser point cloud mapping and matching method and system based on off-road scenes to solve the technical problems of existing technologies such as drift and mismatching in dynamic scenes, lack of semantic understanding, difficulty in distinguishing different types of objects in the environment, and limited performance in unstructured complex environments (such as off-road and forests); feature extraction in the point cloud preprocessing stage consumes a certain amount of computing power and causes a certain amount of delay.

[0004] The present invention provides a laser point cloud mapping and matching method based on off-road scenes, which includes: step 1, collecting off-road environment point clouds, filtering the point clouds to reduce point cloud density, and eliminating dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results to form real-time point cloud data; step 2, performing pose prediction and estimation on vehicle sensor data through Kalman filtering, and correcting the predicted pose to obtain a final pose estimation; step 3, adding the real-time point cloud data with semantic labels to a local incremental dense voxel semantic map through the final pose estimation to achieve incremental update of the local map and form a local incremental dense semantic map; step 4, performing point cloud registration on the real-time point cloud data and the local incremental dense semantic map through a semantic point cloud observation model, and obtaining a point cloud key frame based on the point cloud registration and the final pose estimation; step 5, estimating the accurate pose of the vehicle by minimizing measurement errors and topological constraints using the point cloud key frame and the corresponding pose information, and constructing a consistent map to achieve global consistency optimization of the map and complete the construction of the global semantic map.

[0005] Furthermore, the point cloud semantic segmentation method includes: step 11, sampling, normalizing, convolution, and pooling the original point cloud data; step 12, performing feature learning on the processed point cloud data to extract discriminative feature representations; step 13, based on the feature representation, performing classification prediction on each point through a classifier to obtain the semantic segmentation result of the point cloud data.

[0006] According to a laser point cloud mapping and matching method based on off-road scenes as described in claim 1, it is characterized in that the vehicle sensor includes an inertial measurement unit, a wheel speed meter, and a positioning measurement instrument.

[0007] Furthermore, step 4 also includes: adaptively adjusting the ground point cloud and the constraint factor weights during the registration process between the real-time point cloud and the local incremental dense semantic map to enhance the robustness and accuracy of the matching.

[0008] Furthermore, the method also includes: step 6, after the global semantic map is constructed, the real-time point cloud data collected and generated subsequently is aligned with the global semantic map through the semantic point cloud observation model, and the point cloud key frame is obtained based on the point cloud alignment and the final pose estimation, and then go to step 5 to realize the real-time construction of the global semantic map.

[0009] The present invention also provides a system for laser point cloud mapping and matching based on off-road scenes, the system comprising: a point cloud pre-processing module for collecting off-road environment point clouds, filtering the point clouds to reduce the point cloud density, removing dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results, and forming real-time point cloud data; a posture estimation module connected to the point cloud pre-processing module for performing posture prediction and estimation on vehicle sensor data through Kalman filtering, and correcting the predicted posture to obtain a final posture estimation, and performing point cloud registration between the real-time point cloud data and the local incremental dense semantic map through the semantic point cloud observation model. , according to the point cloud registration and final pose estimation, the point cloud keyframe is obtained; the local incremental dense semantic map construction module is connected to the pose estimation module, and the real-time point cloud data with semantic labels is added to the local incremental dense voxel semantic map through the final pose estimation, so as to realize the incremental update of the local map and form a local incremental dense semantic map; the map optimization construction module is connected to the pose estimation module, which is used to estimate the accurate pose of the vehicle by minimizing the measurement error and topological constraints based on the point cloud keyframe and its corresponding pose information, and construct a consistent map to achieve global consistency optimization of the map and complete the construction of the global semantic map.

[0010] Furthermore, the point cloud processing module is also used to sample, normalize, convolve, and pool the original point cloud data, perform feature learning on the processed point cloud data, extract discriminative feature representations, and classify and predict each point through a classifier based on the feature representation to obtain the semantic segmentation results of the point cloud data.

[0011] Furthermore, the vehicle sensors include an inertial measurement unit, a wheel speed meter, and a positioning measurement instrument.

[0012] Furthermore, the pose estimation module is also used to adaptively adjust the ground point cloud and constraint factor weights during the registration process between the real-time point cloud and the local incremental dense semantic map to enhance the robustness and accuracy of the matching.

[0013] Furthermore, the map optimization construction module is also used to, after the global semantic map is constructed, perform point cloud registration between the subsequently collected and generated real-time point cloud data and the global semantic map through a semantic point cloud observation model to achieve real-time construction of the global semantic map.

[0014] The present invention provides a laser point cloud mapping and matching method and system based on off-road scenarios, which can effectively utilize the inference results of deep neural networks, eliminate dynamic objects, and construct a laser point cloud semantic map. At the same time, it can achieve relatively accurate laser matching and positioning, solving the problems of low real-time mapping in existing technologies, the possibility of false detection, large system computing power requirements, and difficulty in finding a suitable balance between computing power, real-time performance, and algorithm accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flow chart of a laser point cloud mapping and matching method based on off-road scenarios provided by the present invention; Figure 2 Flowchart of the point cloud semantic segmentation method provided by the present invention; Figure 3 A schematic diagram of a laser point cloud mapping and matching system based on off-road scenarios provided by the present invention; Figure 4 Schematic diagram of another laser point cloud mapping and matching system based on off-road scenarios provided by the present invention. DETAILED DESCRIPTION

[0016] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0017] Method Example: The present invention provides a laser point cloud mapping and matching method based on off-road scenes, such as Figure 1 As shown, the method includes the following steps.

[0018] Step 1: Collect off-road environment point clouds, filter them to reduce point cloud density, and remove dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results to form real-time point cloud data; Since existing technologies are sensitive to dynamic objects and are prone to drift and mismatching in dynamic scenes, it is necessary to first access the point cloud semantic segmentation and point cloud target detection results to remove dynamic object point clouds, such as Figure 2 As shown, the point cloud semantic segmentation method includes the following steps.

[0019] Step 11: Sampling, normalization, convolution, and pooling are performed on the original point cloud data; Step 12: Perform feature learning on the processed point cloud data to extract discriminative feature representations; In step 13, based on the feature representation, a classifier is used to perform classification prediction on each point to obtain the semantic segmentation result of the point cloud data.

[0020] Step 2: The vehicle sensor data is subjected to Kalman filtering to perform pose prediction and correction to obtain the final pose estimate. The vehicle sensors include an inertial measurement unit, a wheel speed meter, and a positioning measurement instrument.

[0021] Step 3: Add the real-time point cloud data with semantic labels to the local incremental dense voxel semantic map through the final pose estimation to achieve incremental update of the local map and form a local incremental dense semantic map; Step 4: The real-time point cloud data is aligned with the local incremental dense semantic map through the semantic point cloud observation model. The point cloud keyframe is obtained based on the point cloud registration and final pose estimation. The step 4 also includes: adaptively adjusting the ground point cloud and the constraint factor weights during the registration process between the real-time point cloud and the local incremental dense semantic map to enhance the robustness and accuracy of the matching.

[0022] In step 5, the point cloud keyframes and their corresponding pose information are used to estimate the vehicle's accurate pose by minimizing measurement errors and topological constraints, and a consistent map is constructed to achieve global consistency optimization of the map and complete the construction of the global semantic map.

[0023] Pose graph optimization (Pose Graph Optimization) is a common technique used in the perception and navigation of intelligent devices (such as smart vehicles) to estimate the pose (position and attitude) of the device in the environment and the topology of the map. This optimization method has broad applications in many fields, including smart device navigation, Simultaneous Localization and Mapping (SLAM), autonomous vehicles, and virtual reality. In SLAM, Pose graph optimization is used for simultaneous localization and mapping, optimizing the device trajectory and map structure to achieve accurate environmental modeling. In autonomous vehicles, Pose graph optimization is used to estimate the vehicle's accurate pose for precise localization and path planning. In smart device navigation, Pose graph optimization is used for accurate pose estimation, assisting with navigation and path planning.

[0024] Step 6: After the global semantic map is constructed, the real-time point cloud data collected and generated subsequently is aligned with the global semantic map through the semantic point cloud observation model. Based on the point cloud registration and final pose estimation, the point cloud keyframe is obtained, and then go to step 5 to realize the real-time construction of the global semantic map.

[0025] The present invention provides a laser point cloud mapping and matching method and system based on off-road scenarios, which can effectively utilize the inference results of deep neural networks, eliminate dynamic objects, and construct a laser point cloud semantic map. At the same time, it can achieve relatively accurate laser matching and positioning, solving the problems of low real-time mapping in existing technologies, the possibility of false detection, large system computing power requirements, and difficulty in finding a suitable balance between computing power, real-time performance, and algorithm accuracy and robustness.

[0026] Device Item Example: A system for laser point cloud mapping and matching based on off-road scenes, comprising: a point cloud preprocessing module, a pose estimation module, a local incremental dense semantic map construction module, and a map optimization construction module. Figure 3 shown.

[0027] The point cloud preprocessing module is used to collect off-road environment point clouds, filter the point clouds to reduce the point cloud density, and eliminate dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results to form real-time point cloud data; the point cloud processing module is also used to sample, normalize, convolve, and pool the original point cloud data, perform feature learning on the processed point cloud data, extract discriminative feature representations, and classify and predict each point through a classifier based on the feature representation to obtain the semantic segmentation results of the point cloud data.

[0028] The pose estimation module, connected to the point cloud preprocessing module, is used to perform pose prediction on vehicle sensor data through a Kalman filter, correct the predicted pose, and obtain a final pose estimate. The real-time point cloud data is then aligned with the local incremental dense semantic map using a semantic point cloud observation model. Point cloud keyframes are generated based on the point cloud registration and final pose estimate. The vehicle sensors include an inertial measurement unit, a wheel speedometer, and a positioning measurement instrument. The pose estimation module is also used to adaptively adjust the ground point cloud and constraint factor weights during the alignment process between the real-time point cloud and the local incremental dense semantic map to enhance the robustness and accuracy of the match.

[0029] The local incremental dense semantic map construction module is connected to the pose estimation module. It adds the real-time point cloud data with semantic labels to the local incremental dense voxel semantic map through the final pose estimation, realizes the incremental update of the local map, and forms a local incremental dense semantic map. The map optimization construction module is connected to the pose estimation module and is used to estimate the accurate pose of the vehicle by minimizing the measurement error and topological constraints based on the point cloud keyframes and their corresponding pose information, and to build a consistent map to achieve global map consistency optimization and complete the construction of the global semantic map. The map optimization construction module is also used to, after the global semantic map is built, align the real-time point cloud data collected and generated subsequently with the global semantic map through the semantic point cloud observation model to achieve real-time construction of the global semantic map, such as Figure 4 shown.

[0030] The present invention provides a laser point cloud mapping and matching method and system based on off-road scenarios, which can effectively utilize the inference results of deep neural networks, eliminate dynamic objects, and construct a laser point cloud semantic map. At the same time, it can achieve relatively accurate laser matching and positioning, solving the problems of low real-time mapping in existing technologies, the possibility of false detection, large system computing power requirements, and difficulty in finding a suitable balance between computing power, real-time performance, and algorithm accuracy and robustness.

[0031] In summary, the present invention provides a laser point cloud mapping and matching method and system based on off-road scenarios. This technical solution is different from traditional laser SLAM technology. On the basis of the traditional solution, real-time perceived semantic information is added to the global map. Compared with adding a deep learning model to the traditional SLAM, this solution can efficiently utilize the results of the vehicle's real-time perception module to achieve more robust and accurate matching and positioning results.

[0032] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A laser point cloud mapping and matching method based on off-road scenes, characterized in that: The method comprises: Step 1: Collect off-road environment point clouds, filter them to reduce point cloud density, and remove dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results to form real-time point cloud data; Step 2: The vehicle sensor data is subjected to Kalman filtering to perform pose prediction and correction to obtain the final pose estimate. Step 3: Add the real-time point cloud data with semantic labels to the local incremental dense voxel semantic map through the final pose estimation to achieve incremental update of the local map and form a local incremental dense semantic map; Step 4: The real-time point cloud data is aligned with the local incremental dense semantic map through the semantic point cloud observation model. The point cloud keyframe is obtained based on the point cloud registration and final pose estimation. In step 5, the point cloud keyframes and their corresponding pose information are used to estimate the vehicle's accurate pose by minimizing measurement errors and topological constraints, and a consistent map is constructed to achieve global consistency optimization of the map and complete the construction of the global semantic map.

2. The laser point cloud mapping and matching method based on off-road scenes according to claim 1, characterized in that: The point cloud semantic segmentation method includes: Step 11: Sampling, normalization, convolution, and pooling are performed on the original point cloud data; Step 12: Perform feature learning on the processed point cloud data to extract discriminative feature representations; In step 13, based on the feature representation, a classifier is used to perform classification prediction on each point to obtain the semantic segmentation result of the point cloud data.

3. The laser point cloud mapping and matching method based on off-road scenes according to claim 1, characterized in that: The vehicle sensors include an inertial measurement unit, a wheel speed meter, and a positioning measurement instrument.

4. The laser point cloud mapping and matching method based on off-road scenes according to claim 1, characterized in that: The step 4 also includes: adaptively adjusting the ground point cloud and the constraint factor weights during the registration process between the real-time point cloud and the local incremental dense semantic map to enhance the robustness and accuracy of the matching.

5. The laser point cloud mapping and matching method based on off-road scenes according to claim 1, characterized in that: The method further comprises: Step 6: After the global semantic map is constructed, the real-time point cloud data collected and generated subsequently is aligned with the global semantic map through the semantic point cloud observation model. Based on the point cloud registration and final pose estimation, the point cloud keyframe is obtained, and then go to step 5 to realize the real-time construction of the global semantic map.

6. A system for implementing the off-road scene-based laser point cloud mapping and matching method according to claims 1-5, characterized in that: The system comprises: The point cloud preprocessing module is used to collect off-road environment point clouds, filter the point clouds to reduce the point cloud density, and remove dynamic object point clouds through point cloud semantic segmentation and point cloud target detection results to form real-time point cloud data; The pose estimation module is connected to the point cloud preprocessing module and is used to perform pose prediction on the vehicle sensor data through Kalman filtering, correct the predicted pose, and obtain the final pose estimate. The real-time point cloud data is then aligned with the local incremental dense semantic map through the semantic point cloud observation model. Based on the point cloud registration and the final pose estimate, the point cloud keyframe is obtained. The local incremental dense semantic map construction module is connected to the pose estimation module. It adds the real-time point cloud data with semantic labels to the local incremental dense voxel semantic map through the final pose estimation, realizes the incremental update of the local map, and forms a local incremental dense semantic map. The map optimization construction module is connected to the pose estimation module. It is used to estimate the accurate pose of the vehicle by minimizing the measurement error and topological constraints based on the point cloud keyframes and their corresponding pose information, and to build a consistent map to achieve global consistency optimization of the map and complete the construction of the global semantic map.

7. The laser point cloud mapping and matching system based on off-road scenes according to claim 6, characterized in that: The point cloud processing module is also used to sample, normalize, convolve, and pool the original point cloud data, perform feature learning on the processed point cloud data, extract discriminative feature representations, and classify and predict each point through a classifier based on the feature representation to obtain the semantic segmentation results of the point cloud data.

8. The laser point cloud mapping and matching system based on off-road scenes according to claim 6, characterized in that: The vehicle sensors include an inertial measurement unit, a wheel speed meter, and a positioning measurement instrument.

9. The laser point cloud mapping and matching system based on off-road scenes according to claim 6, characterized in that: The pose estimation module is also used to adaptively adjust the ground point cloud and constraint factor weights during the real-time point cloud and local incremental dense semantic map registration process to enhance the robustness and accuracy of the matching.

10. The laser point cloud mapping and matching system based on off-road scenes according to claim 6, characterized in that: The map optimization and construction module is also used to, after the global semantic map is constructed, perform point cloud registration between the subsequently collected and generated real-time point cloud data and the global semantic map through a semantic point cloud observation model, thereby realizing real-time construction of the global semantic map.

Images