KAM-Net: Keypoint-Aware and Keypoint-Matching Network for Vehicle Detection From 2-D Point Cloud
Tianpei Zou, Guang Chen, Zhijun Li, Wei He, Sanqing Qu, Shangding Gu, Alois Knoll · IEEE Transactions on Artificial Intelligence · 2021
Two-dimesional (2-D) LiDAR is an efficient alternative sensor for vehicle detection, which is one of the most critical tasks in autonomous driving. Compared to the fully developed 3-D LiDAR vehicle detection, 2-D LiDAR vehicle detection has much room to improve. Most existing state-of-the-art works represent 2-D point clouds as pseudo-images and then perform detection with traditional object detectors on 2-D images. However, they ignore the sparse representation and geometric information of vehicles in the 2-D cloud points. To address these issues, in this article, we present a novelkeypoint-aware and keypoint-matching networktermed as KAM-Net, which focuses on better detecting the vehicles by explicitly capturing and extracting the sparse information of L-shape in 2-D LiDAR point clouds. The whole framework consists of two stages—namely, keypoint-aware stage and keypoint-matching stage. The keypoint-aware stage utilizes the heatmap and edge extraction module to simultaneously predict the position of L-shaped keypoints and inflection offset of L-shaped endpoints. The keypoint-matching stage is followed to group the keypoints and produce the oriented bounding boxes with axis by utilizing the endpoint-matching and L-shaped-matching methods. Further, we conduct extensive experiments on a recently released public dataset to evaluate the effectiveness of our approach. The results show that our KAM-Net achieves a new state-of-the-art performance. The source code is available athttps://github.com/ispc-lab/KAM-Net.