Keypoint3D: Keypoint-Based and Anchor-Free 3D Object Detection for Autonomous Driving with Monocular Vision

Autonomous driving has received enormous attention from the academic and industrial communities. However, achieving full autonomy is not a trivial task, because of complex dynamic environment. Perception ability tough challenge for autonomous driving, while 3D object detection serves as breakthrough providing precise dependable geometric information. Inspired by practical experiences human experts, pure visual scheme takes sufficient responsibility safe stable driving. In this paper, we proposed an anchor-free keypoint-based detector with monocular vision, named Keypoint3D. We creatively leveraged 2D projected points objects’ centers keypoints modeling. Additionally, positioning, utilized novel self-adapting ellipse Gaussian filter (saEGF) on heatmaps, considering different shapes. tried variations DLA-34 backbone semi-aggregation (SADLA-34) network, which pruned redundant aggregation branch but achieved better performance. Keypoint3D regressed yaw angle in Euclidean space, resulted closed mathematical space avoiding singularities. Numerous experiments KITTI dataset moderate level have proven that best speed-accuracy trade-off average precision 39.1% at 18.9 FPS cars detection.