监控视频下的多尺度车辆检测与跟踪  

Multi-scale vehicle detection and tracking under surveillance video

作　　者：肖永超 吴训成 刘康安 朱肖磊 XIAO Yongchao;WU Xuncheng;LIU Kang'an;ZHU Xiaolei(Shanghai University of Engineering and Technology,Mechanical and Automotive Engineering,Shanghai 201620,China)

机构地区：[1]上海工程技术大学机械与汽车工程学院,上海201620

出　　处：《激光杂志》2022年第1期39-44,共6页Laser Journal

摘　　要：车辆检测与跟踪在自动驾驶系统中起着重要的作用。针对交通监控视频中多尺度车辆目标难以检测和重叠目标容易漏检等问题,提出使用倒残差方法来改进YOLOv3的卷积层提取特征,在YOLOv3网络中增加空间金字塔(SPP)模块,获取车辆多尺度信息,并采用Soft-NMS代替非最大值抑制(NMS),减少车辆重叠导致的预测框丢失。最后使用KCF算法对目标车辆进行跟踪。在KITTI数据集上的实验结果表明,该方法对不同尺度的车辆目标都取得了较好的检测跟踪效果,能更好地满足实际应用的需要。Vehicle detection and tracking play an important role in autonomous driving systems. Aiming at the problems that multi-scale vehicle targets are difficult to detect and overlapping targets are easily missed in traffic surveillance videos, an inverted residual method is proposed to improve the convolutional layer extraction features of YOLOv3, add a spatial pyramid(SPP) module to the YOLOv3 network to obtain vehicle multi-scale information, and use Soft-NMS instead of non-maximum suppression(NMS) to reduce the prediction box caused by overlapping vehicles is lost. Finally, the KCF algorithm is used to track the target vehicle. The experimental results on the KITTI data set show that this method has achieved better detection and tracking results for vehicle targets of different scales, and can better meet the needs of practical applications.

关 键 词：多尺度 YOLOv3 目标检测 目标跟踪 KITTI 

分 类 号：TN209[电子电信—物理电子学]