基于改进PointPillars的3D目标检测算法
Research on 3D target detection algorithm based on PointPillars
谢生龙 1邵金菊 1单少飞 1孙福昌 1王磊2
作者信息
- 1. 山东理工大学交通与车辆工程学院,山东淄博 255000
- 2. 睿羿科技(山东)有限公司,山东淄博 255000
- 折叠
摘要
针对自动驾驶场景下远距离及遮挡目标识别问题,对PointPillars算法进行了改进.引入了并行的空间注意力和通道注意力机制,增强了目标的位置信息及有用特征通道权重,提高了远距离目标的检测精度.在2D CNN骨干网络中引入了自适应空间特征融合模块,解决了特征拼接的信息丢失问题,提高了遮挡目标的检测精度.基于KITTI数据集在3种不同场景难度下分别对SECOND、PointPillars、改进PointPillars这3种算法进行了定量分析验证,并将改进的PointPillars算法进行可视化分析.定量分析表明,改进PointPillars算法在鸟瞰图模式下目标检测精度最大提升2.75％;在三维模式下目标检测精度最大提升2.93％;在AOS模式下目标检测精度最大提升4.05％,可视化结果表明改进的PointPillars算法能有效检测远距离及遮挡目标.
Abstract
To better identify long-distance and occluded targets in autonomous driving scenarios,this paper improves the PointPillars algorithm.Parallel spatial attention and channel attention mechanisms are introduced to enhance the target's position information and useful feature channel weights and thus improve the detection accuracy of long-distance targets.An adaptive spatial feature fusion module is employed in the 2D CNN backbone network,which addresses information loss in feature splicing and improves the detection accuracy of occluded targets.Based on the KITTI dataset,this paper conducts quantitative analysis and verification on SECOND,PointPillars,and our improved PointPillars algorithm under three different scenarioes.Our improved PointPillars algorithm is also visualized.Results show our method achieves a maximum improvement of 2.75％in target detection accuracy in bird's-eye view mode,of 2.93％in three-dimensional mode,and of 4.05％in AOS mode.Our visualization results indicate the improved PointPillars algorithm effectively detects long-distance and occluded targets.
关键词
目标检测/PointPillars/注意力机制/点云/自适应空间特征融合Key words
target detection/PointPillars/attention mechanism/point cloud/adaptive spatial feature fu-sion引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据