OC-BEVDepth: An Object Contour-Perceiving 3D BEV Object Detection Framework

In autonomous driving systems, 3D object detection plays a pivotal role. In spite of the fact that lidar-based methods have been proven to be capable of providing reliable 3D detection results, multi-view camera-based methods are attracting more and more attention as a result of their lower cost. Significant strides have been achieved in the development of 3D object detection using the Lift-Splat-Shot (LSS) approach. Despite advancements, imprecise depth estimation remains a critical factor constraining the precision of purely visually-driven 3D target detection models, especially in the foreground target edge area, where significant depth variations occur, known as the “depth mutation” challenge. In this paper, a new object contour sensing framework (OC-BEVDepth) is proposed. Firstly, an Object Contour Depth Fusion (OCDF) module is developed to solve “depth mutation” problem. Secondly, the previous image depth prediction module is optimized by Dense Depth Refinement (DDR) module to perform depth supervision in the input dimension. Experiments conducted on nuscenes benchmarks confirm the effectiveness of our OC-BEVDepth approach. It achieved mAP and NDS of 41.4% and 42.1 %, respectively, on the nuScenes benchmark, outperforming BEVDepth by 10.7% and 5.4%, respectively. Additionally, the feasibility of our method is validated on a custom dataset.