基于改进YOLOv5的三维目标检测算法  

3D Object Detection Algorithm Based on Improved YOLOv5

作　　者：盛雪清 李绍斌[1] 屈金燕 刘留[1] Sheng Xueqing;Li Shaobin;Qu Jinyan;Liu Liu(School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学电子信息工程学院,北京100044

出　　处：《激光与光电子学进展》2024年第18期137-147,共11页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(52102472)。

摘　　要：针对三维目标检测点云数据量大且对小目标物体检测识别效果较差的问题,基于Complex-YOLO算法思想,提出了一种改进YOLOv5网络的点云三维目标检测方法。首先,针对点云数据量大导致后续网络运行时间过长的问题,采用Complex-YOLO算法,将点云数据转化为RGB-Map格式,易于YOLOv5网络处理,同时给YOLOv5增加了角度预测分支和旋转框回归损失计算方法,用于实现对RGB-Map中旋转目标的精确定位。其次,为了提升网络对小目标的感知能力,对YOLOv5网络结构进行改进,引入了小目标特征融合层和预测头用于增强算法对小目标物体的检测能力。最后,在neck网络中增设卷积块注意力模块(CBAM),提升网络对小目标的敏感程度。在KITTI数据集上进行验证,实验结果表明,本文提出的基于改进YOLOv5网络的三维目标检测方法相较于Complex-YOLO算法,Car类型的平均精度均值(mAP)提高了7.48百分点,Pedestrian类型的mAP提高了12.54百分点,Cyclist类型的mAP提高了1.2百分点,所有类型的mAP提高了7.08百分点,证明了本文算法的有效性。To address the challenge of handling large volumes of point cloud data for three-dimensional(3D)object detection and the limited effectiveness in detecting small objects,in this study,an enhanced 3D target detection method is proposed that improves the YOLOv5 network based on the idea of Complex-YOLO algorithm.The proposed approach first tackles the issue of lengthy processing times due to extensive point cloud data by adopting the Complex-YOLO strategy of converting point cloud data into an RGB-Map format,which is more manageable for the YOLOv5 network.Enhancements to YOLOv5 include an angle prediction branch and a rotation frame regression loss function to accurately position rotating targets within the RGB-Map.Additionally,the YOLOv5 architecture is modified to better detect small objects by incorporating a feature fusion layer and a dedicated prediction head,which heightens the network’s sensitivity to smaller targets.Furthermore,the convolutional block attention module(CBAM)attention mechanism is integrated into the network’s neck to further enhance detection sensitivity.Experimental evaluations on the KITTI dataset confirm the superiority of the modified YOLOv5 method over the original Complex-YOLO,with improvements in mean average precision(mAP):Car type mAP increased by 7.48 percentage points,Pedestrian type by 12.54 percentage points,Cyclist type by 1.2 percentage points,and an overall increase of 7.08 percentage points across all categories,demonstrating the effectiveness of this algorithm.

关 键 词：三维目标检测 YOLOv5 Complex-YOLO 小目标检测 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]