EXTRACTING VEHICLES IN POINT CLOUDS OF UNDERGROUND PARKING LOTS BASED ON GRAPH CONVOLUTION

Three-dimensional point clouds can describe the shape and position of objects more accurately when compared with 2D images, thereby providing richer information for object recognition, detection, and reconstruction tasks. Extracting vehicles in point clouds of underground parking lots can help autonomous vehicles achieve automatic parking. Camera-based perception algorithms will fail in complicate environments, so it is necessary to study algorithms for extracting targets using point cloud data. In this paper, we designed an effective method to extract the vehicles in the underground parking lot. First, the point clouds belonging to the vehicle will be segmented using a neural network based on graph convolution, and then different vehicles will be separated based on clustering. Finally, the minimum bounding box for each car is calculated. The proposed approach achieved much better results on the point cloud dataset than other state-of-the-art methods. Our method achieves 99.6% in Overall Accuracy and 98.5% in Mean IOU (Intersection over Union).