基于光电转台全局运动补偿与YOLOv3网络模型的运动目标检测算法  

Moving Target Detection Based on Global Motion Compensation of Photoelectric Turntable and YOLOv3 Network Mode

作　　者：马志扬 傅慧妮 赵博[1] 杨忠琳 姜雨彤 朱梦琪 孙伟琛 李兴鑫 MA Zhiyang;FU Huini;ZHAO Bo;YANG Zhonglin;JIANG Yutong;ZHU Mengqi;SUN Weichen;LI Xingxin(China North Vehicle Research Institute, Beijing 100072, China)

机构地区：[1]中国北方车辆研究所,北京100072

出　　处：《车辆与动力技术》2022年第1期52-59,34,共9页Vehicle & Power Technology

摘　　要：传统的动目标检测算法存在受噪声干扰大、背景建模时间长、无法识别目标类型等问题,并且背景建模等方法对全局运动存在限制要求.针对上述问题,提出新的运动目标检测算法,采用YOLOv3模型对观察序列中的视频进行检测,获得连续帧的目标识别结果;通过光电转台位置信息估算全局运动,对连续帧目标进行匹配获取一致性关系;计算目标像素速度与目标运动程度估值,判定目标运动状态.以车辆目标视频图像为研究对象,试验表明:算法在英伟达TX2与NX平台实现部署,单帧处理时间分别达到140 ms、36 ms,实现对多个运动目标的实时正确检测,并获取目标像素速度与运动方向等运动信息.The problems of moving target detection by traditional means include large noise interference,long background modeling time,and unable to identify the type of targets.Global motion is limited using modelling methods such as background model.To solve the above problems,a new moving target detection algorithm is proposed.The key approach is firstly to detect the video in the observation sequence to obtain the target recognition results by using YOLOv3.Then the global motion is estimated by the position information of the photoelectric turntable to match targets in the continuous frame.Finally,the target moving state is determined by calculating the target pixel speed and motion degree estimation.The algorithm is deployed on NVIDIA TX2 and NX platforms to test the vehicle videos.The processing time of single frame reaches 140 ms and 36 ms respectively.Multiple moving targets are detected correctly in real time,and the motion information such as target pixel speed and moving direction are obtained.

关 键 词：YOLOv3 全局运动补偿 动目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]