基于改进YOLOv7的自动驾驶目标检测方法
Automatic driving target detection method based on improved YOLOv7
程换新 1徐皓天 1骆晓玲1
作者信息
摘要
针对自动驾驶场景下,车辆目标密集、相互遮挡和目标过小导致的误检、漏检问题,提出一种改进YOLOv7的车辆目标检测算法.在主干网络SPPCSPC后加入ACmix混合注意力机制,充分挖掘特征信息,增强网络对车辆信息的关注度,减少其他目标的干扰,提高检测精度;在Neck端中加入Swin Transformer,收集全局信息;添加160x160尺寸目标检测头,以增加锚点的数量和密度,提高网络对小目标的感知能力;最后利用Soft-NMS柔性非极大值抑制剔除冗余候选框,改善漏检能力.通过实验验证了改进的可行性并与五种主流网络进行了对比,平均精度达到91.5％,与基础网络YOLOv7相比,平均精度提高7.1％,运行速度达到105 FPS,证明了改进方法的有效性.
Abstract
Aiming at the problem of misdetection and omission caused by dense vehicle targets,mutual occlusion and too small targets in automatic driving scenarios,a vehicle target detection algorithm with improved YOLOv7 is pro-posed.The ACmix hybrid attention mechanism is added after the SPPCSPC of the backbone network to fully mine the feature information,enhance the network's attention to the vehicle information,reduce the interference of other targets,and improve the detection accuracy;the Swin Transformer is added to the Neck end to collect the global information;the 160x160-size target detection head is added to increase the number and density of the anchors and to improve the network's ability to perceive small targets;finally,Soft-NMS flexible non-maximum suppression is utilized to reject redundant candidate frames and improve the leakage detection ability.The feasibility of the improvement is verified by experiments and compared with five mainstream networks,and the average accuracy reaches 91.5％,and compared with the basic network YOLOv7,the average accuracy is improved by 7.1％,and the operation speed reaches 105 FPS,which proves the effectiveness of the improved method.
关键词
自动驾驶/目标检测/YOLOv7/ACmix/Swin/TransformerKey words
automatic driving/object detection/YOLOv7/ACmix/Swin Transformer引用本文复制引用
出版年
2024
NETL
NSTL
万方数据