基于改进PV-RCNN的3D目标检测算法实验研究

Experimental research on 3D target detection algorithm based on improved PV-RCNN

3D目标检测结合了深度信息,能够提供目标的位置、方向和大小等空间场景信息,在自动驾驶和机器人领域发展迅速。针对PV-RCNN在3D目标检测时不能够充分适应不同的物体尺度、不同的点云密度、部分变形和杂波等问题,对3D目标检测的任务进行实验研究。通过加入自适应可变形卷积、上下文融合模块和Gumbel Subset Sampling模块来训练层级特征,使得编码关键点自适应地朝着最具有判别和代表性的特征对齐,提高提案框回归精度。实验结果表明,改进后的PV-RCNN 3D目标检测精度得到了提升,尤其是在远距离物体识别和检测方面。

3D target detection combines depth information to provide spatial scene information such as the position,direction, and size of the target, and is developing rapidly in the field of autonomous driving and robotics. Aiming at the problems that PV-RCNN can not fully adapt to different object scales, different point cloud densities, partial deformations and clutter in 3D target detection, experimental research on the task of 3D target detection is carried out. By adding adaptive deformable convolution, context fusion module and Gumbel Subset Sampling module to train hierarchical features, the coding key points are adaptively aligned to the most discriminative and representative features, and the regression accuracy of proposal box regression is improved. The experimental results show that the improved PV-RCNN 3D target detection accuracy has been improved, especially in long-distance object recognition and detection.