Abstract: The traditional climbing detection method is based on monocular vision,which has specific limit and low accuracy.In order to detect the climbing behavior more accurately in video surveillance,this paper proposes a climbing detection method based on binocular stereo vision.The feature points of both rectified images are extracted and the vision disparity of the point is obtained by the sparse stereo matching method.The 3D coordinates of the points are calculated based on the standard 3D measurement theory.Targets are detected by points clustering.According to targets’ location and modified joint probabilistic data association method,targets can be tracked.Based on targets’ location,the trajectory and moving direction are analyzed to detect the behavior of climbing.Experimental result shows that the method has high accuracy in the conditions of complex background and the average accuracy is 94% with appropriate parameters.

Key words: climbing detection, sparse stereo matching, spatial point clustering, object tracking, motion trajectory

摘要： 传统基于单目视觉的攀爬检测大多难以解决光线变化及阴影干扰等问题,具有一定局限性,可靠性较低。为在视频监控中准确地检测攀爬行为,提出一种新的攀爬检测方法。该方法利用极线校正后的左右图像,通过稀疏立体匹配方法计算得到匹配点和对应的视差值,并使用标准三维测量原理求取特征点的三维坐标,对离散特征点采用空间点聚类方法检测目标,以俯视图中目标的位置为轨迹点,结合修正的联合概率数据关联方法进行跟踪,根据目标的位置分析运动轨迹和运动方向,从而检测出目标是否有攀爬行为。实验结果表明,该方法在复杂背景下检测攀爬行为的准确率较高,当参数大于等于0.4时平均准确率约为94%。

关键词: 攀爬检测, 稀疏立体匹配, 空间点聚类, 目标跟踪, 运动轨迹

CLC Number: 

• TP391