6DoF-3D: Efficient and accurate 3D object detection using six degrees-of-freedom for autonomous driving

Abstract

With the advancement of automobile industry, autonomous driving has become an integral component of contemporary vehicles. In this context, the 3D spatial perception of vehicles plays a crucial role. The complexity of 3D object detection dramatically increased due to the explosion of degrees-of-freedom for 3D bounding boxes. Therefore, current 3D detection algorithms struggle to achieve high accuracy and efficiency simultaneously. To address this problem, we introduce a single-stage, LiDAR-based, and efficiency–accuracy balanced 3D object detector named 6DoF-3D. The proposed network architecture comprises a backbone of CSPDarknet53, a multi-scale feature fusion neck, and a detection head. We transform 3D point clouds into pseudo-images, with critical spatial features retained. 6DoF-3D presents a novel encoding strategy to prune the redundant degree-of-freedom in 3D detection, resulting in a lightweight network architecture. We put forward a 2.5D Euler-Region-Proposal network for precise classification and regression of 2.5D bounding boxes. The pruned dimension is compensated with a geometric transformation decoder. Extensive experiments conducted on the KITTI dataset demonstrate that 6DoF-3D achieves state-of-the-art accuracy of 89.62% for 3D car detection at the moderate level. Additionally, the mean average precision reaches 66.72% at the while maintaining a real-time performance of 33.21 FPS for cars, pedestrians, and cyclists.