Objective In the field of autopilot, the intelligent driving decision is highly dependent on the accurate detection of road obstacles, and the accurate information of road obstacles is the prerequisite for automatic driving. Three dimensional (3D) LiDAR has a high resolution, high detection accuracy, all-weather operation, and strong anti-jamming ability, allowing it to provide accurate environmental information for vehicles. To detect road obstacles using a 3D point cloud generated by 3D LiDAR, the point cloud data must be processed. The point cloud is divided into independent subsets, each of which corresponds to targets with distinct physical meanings. The target detection of a point cloud is divided into several modules, the most important of which are preprocessing, ground segmentation, point cloud clustering, target output, and other modules that lack the target’s pose information. Pose information is critical for tracking, classification, and other tasks involving target obstacles, so it is critical to accurately detect the target’s pose.
Methods The target detection based on 3D point cloud is generally divided into preprocessing, ground segmentation, point cloud clustering, target output, and other modules. On this basis, the target’s pose information is determined further. A segmentation method based on a fan-shaped box is proposed to segment the ground point cloud and the target point cloud to meet the requirements of high accuracy and high real-time for ground point cloud segmentation. To address the issue of Euclidean clustering’s fixed distance threshold’s poor adaptability to the characteristics of LiDAR close dense far sparse, a Euclidean clustering method with adaptive clustering threshold is proposed to improve the clustering accuracy of the clustering algorithm at different distances. A boundary line fitting method based on the random sample consensus (RANSAC) algorithm is proposed to determine the target object’s box model to meet the requirement of pose information detection. The direction angle is obtained by direction fitting, and the dimension is obtained by coordinate system rotation, which reduces the amount of computation and enhances the real-time performance of the algorithm. RANSAC direction and measurement estimation is the name of the pose detection algorithm used in this paper (RDME).To validate the algorithm, 100 frames point clouds are randomly selected from each of the three LiDAR datasets during the experiments. Statistics of the number of objects in all single frame point clouds and the number of targets detected by the clustering algorithm and pose detection algorithm record the processing time of each frame and use four indicators to evaluate the algorithm effect, and the algorithm is compared with the minimum bounding rectangle (MBR) algorithm and the Three-Point Estimation (TPE) algorithm to verify the algorithm’s effectiveness and advanced nature.
Results and Discussions The clustering algorithm performs well, with high true positive accuracy (TPA) and low false negative accuracy (FNA). RDME algorithm performs better than MBR and TPE algorithms in three LiDAR datasets. The RDME algorithm performs better when the number of LiDAR laser rays is increased and the point cloud is dense, and the total true positive accuracy (TTPA) is improved by up to 4.52 percentage points. For PPA, the performance of RDME algorithm is improved by up to 8.14 percentage points. The RDME algorithm can effectively estimate the direction of the target object, and the adaptive threshold clustering method based on LiDAR parameters can effectively identify the target objects at different distances and improve the clustering algorithm’s accuracy (Table 2). In the same frame sequence, the detection speeds of the RDME, MBR, and TPE algorithms differ. The average detection time of the RDME algorithm is 14.84 ms, 19.48 ms, and 19.71 ms respectively (excluding the link time of data transmission) in the first LiDAR datasets, and RDME algorithm detection time is 23.81% lower than the MBR algorithm, and 24.71% less than TPE algorithm; the average detection time of the RDME algorithm is 38.10 ms, 47.02 ms and 48.07 ms respectively in second LiDAR datasets, and the RDME algorithm detection time is 18.97% lower than MBR algorithm and 20.74% less than TPE algorithm; the average detection time of the RDME algorithm is 12.78 ms, 15.3 5 ms and 15.56 ms respectively in the third LiDAR datasets, and the RDME algorithm detection time is 16.74% lower than MBR algorithm and 17.86% less than TPE algorithm (Fig. 11). The results show that the RDME algorithm has less detection time and better real-time performance than the MBR algorithm and TPE algorithm.
Conclusions A ground segmentation method based on sector bin box and slope threshold is proposed, which can meet the real-time and accuracy requirements of ground point cloud segmentation. This paper proposes an algorithm for determining the clustering threshold based on LiDAR performance parameters to process the point cloud, overcoming the problem that European clustering relies on setting the distance threshold. For obstacles at different distances, the clustering algorithm can achieve a better clustering effect. A pose detection (RDME) algorithm is proposed to detect the pose information of the target obstacle based on the feature that the point cloud is concentrated on the surface of the target object after projection. The position detection (RDME) algorithm is compared to the MBR algorithm and the TPE algorithm using three different types of LiDAR datasets. Compared with the MBR algorithm, the RDME algorithm improves detection accuracy by 1.12%, 3.98%, and 8.14%, and reduces detection time by 23.81%, 18.97%, and 16.74%; compared with the TPE algorithm, the detection accuracy of the RDME algorithm is improved by 2.43%, 3.63%, and 6.57% respectively, and the detection time is reduced by 24.71%, 20.74%, and 17.86% respectively. The RDME algorithm’s detection accuracy and real-time performance have been significantly improved. The algorithm’s maximum average detection time is 47.02 ms, which is less than 0.1 s of the 10 Hz LiDAR’s real-time requirement. It can meet the real-time needs of intelligent vehicles and is a more efficient method.