This study aims to enhance 3D object detection in urban traffic scenarios by addressing the limitations of bird’s eye view (BEV) generation using the lift-splat-shoot (LSS) framework, particularly feature loss in occluded regions and depth estimation instability, which hinder safe autonomous navigation.
An improved LSS algorithm was developed to better handle depth estimation and feature alignment. It incorporates an adaptive probability transition matrix that extends depth distributions in occluded areas. It also uses a cosine annealing strategy to optimize depth denoising. An instance alignment module is introduced to align BEV and instance features. In addition, a combined loss function based on uncertainty sampling is used to balance semantic consistency. The approach was validated through experiments on the nuScenes data set, comparing performance against the traditional LSS algorithm and a baseline method.
The proposed algorithm achieved a comprehensive score (NDS) of 0.514 and a mean average precision (mAP) of 0.422, outperforming the traditional LSS algorithm (NDS: 0.436, mAP: 0.330) and the baseline method (NDS: 0.475, mAP: 0.351). It significantly improved target detection accuracy and depth estimation stability in occlusion scenarios, enhancing overall perception reliability.
This research introduces a novel combination of adaptive probability transition and cosine annealing strategies within the LSS framework, substantially improving BEV generation robustness. The findings benefit autonomous driving systems by enabling more accurate 3D object detection in complex urban environments, supporting safer navigation.
