摘要
针对周界视频监控应用环境特殊的问题,提出了一种人员翻越行为的检测方法。该方法采用"目标检测-人员跟踪-轨迹分析"的流程。在算法的人员跟踪过程中,将混合高斯模型得到的前景区域与KLT光流法得到的特征点运动信息结合起来,提出了一种新型跟踪算法。该算法仅使用图像的灰度信息作为输入,一定程度上能够适应目标形变及遮挡,并具有很强的鲁棒性和实时性;在算法的轨迹分析过程中,结合墙体位置信息与先验知识设计了一种新的轨迹分析的方法,不需要通过在线学习可直接对轨迹进行分析。实验结果表明,该算法在测试视频集上检测准确率超过93%,与现有方法相比,能更好地适应实际应用中复杂的环境条件。
An algorithm to detect fence climbing is proposed to meet the specific requirements of perimeter video surveillance.The algorithm bases on a process like "object detection-trackingtrajectory analysis".The main scheme of the algorithm combines foreground areas obtained from Gaussian mixture model with displacements of feature points obtained from KLT algorithm in personnel tracking stage.In this way,the tracking algorithm just takes advantage of the image grey scale information,can adapt deformation and shade of the objects in a certain extent and has a strong robustness as well as real-time effect.A new method for the analysis of trajectory is proposed in trajectory analysis stage by using the information of wall location and other prior knowledge instead of traditional online learning algorithms.Experimental results show that the proposed algorithm is more suitable for noisy environments than other existing algorithms.A recognition rate above 93%is obtained on a test video set.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2016年第6期47-53,共7页
Journal of Xi'an Jiaotong University
基金
天津市科技支撑计划重点资助项目(14ZCZDSF00020)
关键词
视频监控
异常检测
跟踪
轨迹分析
video surveillance
anomaly detection
tracking
trajectory analysis