Adaptive Threshold Clustering Segmentation Method Based on Two-Dimensional Lidar
Wang Zhu,Wang Zhi,Zhang Xu,Cui Can,Wang Jian
DOI: https://doi.org/10.3788/CJL202118.1610005
2021-01-01
Abstract:Objective With the continuous advancement of science and technology, research into driverless smart cars has increased significantly. To realize environmental perception, driverless smart cars rely on sensors installed in the platform to collect and process surrounding environmental information. Sensors that perceive the external environment primarily include light detection and ranging (lidar), millimeter-wave radar, ultrasound, and cameras. Lidar is widely used in smart car obstacle detection and mapping navigation systems due to its high accuracy, large detection range, anti-interference properties, and the degree of depth information. Currently, two-dimensional lidar is widely used for obstacle detection. Obstacle clustering segmentation is a key technology in environment perception. Due to the complex environment characteristics and uneven distribution of data density, traditional clustering algorithms cannot achieve good clustering of different types of obstacles at different distances simultaneously, and obstacles can be missed and misdetected. Therefore, considering indoor obstacle detection requirements, the characteristics of lidar data and the geometric characteristics of indoor environments are analyzed and an improved adaptive threshold algorithm based on distance and obstacle characteristics is proposed. With the proposed algorithm, the threshold value is adjusted to change with the target distance and intra-class density. This method can effectively improve the correct detection rate of obstacles, and the segmentation accuracy reaches 92. 23%. Methods The proposed algorithm introduces the cluster density concept of the DBSCAN algorithm and an improved linear threshold algorithm. First, after receiving lidar data, the upper computer preprocesses the data, eliminates unstable points outside the effective range, and converts the remaining points to a rectangular coordinate system. Then, distance and density thresholds are set. The distance threshold changes adaptively with the distance of the received lidar data. The density threshold is the density average of the current class plus a multiple of its density standard deviation, which can vary as the obstacle class inner density changes adaptively. Then, the starting point is classified as category 1. From the second point to the last point, it is determined whether the distance between the current point and the previous point is greater than the current distance threshold. If the distance is less than the threshold, it is classified into the current category. Otherwise, a new class is created, and it is classified into the new class. It is also determined whether the last point and the starting point are less than the threshold. If less, the last category will be classified into the first category, otherwise it will not be included. Results and Discussions The indoor test results are shown in Fig. 6. As shown in Fig. 6( b), it is difficult for the linear threshold algorithm to consider the sparse scanning points of distant obstacles and obstacles closely distributed at a short distance. It can be seen from Fig. 6(c) that the improved DBSCAN algorithm will group different obstacles that are close together into one category. In Fig. 6 ( d), in the vicinity of the same coordinates, the proposed adaptive threshold algorithm based on distance and obstacle characteristics can distinguish smaller object distances at the same distance and can successfully segment objects with different distance data densities. The outdoor test results are shown in Fig. 8. Under complex outdoor conditions, the proposed, DBSCAN, and linear threshold algorithms successfully segmented and detected various obstacles at close distances. However, the linear threshold method incorrectly classified walls as noise, and the improved DBSCAN algorithm mistakenly detected walls as two separate classes. For metal aluminum plates, the linear threshold algorithm and the improved DBSCAN algorithm mistakenly detected several types of obstacles. The proposed adaptive threshold algorithm successfully segmented and detected all types of obstacles that are close to each other and did not cause over-segmentation of walls and aluminum plates at a distance. In addition, the positive detection rate is significantly improved. Conclusions The experimental results showed that the improved clustering segmentation algorithm had the following advantages. First, it improved the positive detection rate of obstacle clustering and segmentation. Compared with the linear threshold algorithm, it had stronger environmental adaptability and could adapt to longer environmental distances and smaller object intervals. Second, compared with the global search of the DBSCAN algorithm, run time is reduced and the efficiency of obstacle detection and segmentation is improved.