Fast and Accurate: The Perception System of a Formula Student Driverless Car

To detect the cone of the Formula Student Autonomous China (FSAC) track quickly and accurately, this paper proposed a fast and accurate perception system that fused the 3D LIDAR point cloud data and visual information. First, the vision part spliced the images of the two monocular cameras vertically and adopted the YOLOv3 target detection network with the attention mechanism to obtain the type of candidate targets and their position on the pixel plane. Afterward, the track boundary could be extracted and the drivable area could be segmented using cubic polynomial regression by the center of the bounding box. The LIDAR pipeline used the GPF algorithm and the Euclidean clustering algorithm to process the original point cloud to obtain the three-dimensional information of the cones’ coordinates. Finally, the 3D LIDAR point cloud data of reconstructed cones was projected onto the two-dimensional pixel plane and whether the target was detected was determined by judging whether the projected point was in a specific position within the bounding box. Experimental results showed that the proposed method had a larger perception range, faster detection speed, and higher detection accuracy compared with the 2020 season's perception system.