摘要
监控视频的多目标跟踪是视频智能分析的热点研究内容,其中目标的检测是目标跟踪的基础,精度高、速度快的目标检测器对于后续的实时分析任务尤为重要。提出一种针对监控视频的基于差分筛选的YOLOv2目标检测算法,采用差分算法筛选无前景目标帧及设置重叠度量阈值进行跨帧检测,改善了YOLOv2作为检测方法用于监控视频多目标跟踪任务时速度过慢的不足,同时高精度的检测结果有利于下一步多目标跟踪任务的顺利完成。利用NPLR监控视频数据集对YOLOv2目标检测算法进行了测试,并将该方法与可变型部件模型DPM进行了比较。结果表明,差分YOLOv2方法在精度上高出DPM方法0.304 6,检测时间快了26倍左右,验证了该算法的有效性。
Multi-object tracking of surveillance video is a hot research topic of video intelligence analysis,and the detection of the object is the basis of the object tracking.High-precision and fast object detectors are especially important for subsequent real-time analysis tasks.The YOLOv2 object detection algorithm is proposed which based on differential filter for surveillance video.The difference algorithm is used to screen the foreground frames without foreground and the overlap metric threshold is set for cross-frame detection.The shortcomings of YOLOv2 as a detection method for monitoring video multi-target tracking tasks is improved,and the high-precision detection results are beneficial to the next step that the multi-object tracking task is easily completed.The YOLOv2 target detection algorithm is tested using NPLR surveillance video dataset and compared with the deformable parts models (DPM).The results show the accuracy of the differential YOLOv2 method is 0.304 6 higher than that of DPM method and the detection time is about 26 times faster,which verify the effectiveness of the algorithm.
作者
张旭
李建胜
郝向阳
程相博
李朋月
ZHANG Xu;LI Jiansheng;HAO Xiangyang;CHENG Xiangbo;LI Pengyue(Information Engineering University,Zhengzhou 450001,China)
出处
《测绘科学技术学报》
CSCD
北大核心
2018年第6期616-621,共6页
Journal of Geomatics Science and Technology
基金
国家自然科学基金项目(61173077)
国家高技术研究发展计划项目(2015AA7034057A)
关键词
目标检测
多目标跟踪
监控视频
可变型部件模型
深度神经网络
object detection
multi-object tracking
surveillance video
deformable parts models
deep neural network
