Enhanced PV-RCNN Algorithm for 3D Object Detection
Kaixuan He, Xiaoru Song, Yilan Wang, Meng M. Zhao · 2024
3D object detection, leveraging depth information to deliver spatial attributes of targets including location, orientation, and size, is advancing rapidly in autonomous driving and robotics. Addressing challenges faced by PV-RCNN in 3D scenarios such as inefficient sampling, information loss during conversion, and limitations in local feature representation, this paper introduces systematic optimization strategies. Enhanced foreground recognition and background noise suppression are achieved through semantic-guided sampling. Improved sampling efficiency is realized by substituting voxel query for ball query. Attention mechanisms are incorporated to refine detail feature perception. Extensive experiments on standard 3D object detection benchmarks validate the efficacy of the proposed optimizations, demonstrating significant performance enhancements in PV-RCNN across multiple evaluation metrics.