联合加权重构轨迹与直方图熵的异常行为检测

Abnormal behavior detection of joint weighted reconstruction trajectory and histogram entropy

为解决多目标打斗、抢劫等异常行为检测精度不高的问题,提出一种联合加权重构轨迹与直方图熵的异常行为检测算法。首先,采用背景相减法结合宽高比提取行人目标;然后将卡尔曼滤波器及HOG特征融入时空上下文算法中,实现短时间内被完全遮挡行人的鲁棒跟踪;最后对跟踪轨迹进行训练,构造正常行为字典并稀疏重构待检测轨迹,通过联合加权最小重构残差和直方图熵,实现对异常行为的有效检测。通过对比实验,表明该算法对于打斗和抢劫等异常行为具有较好的检测效果,在静态背景且无遮挡的情况下,检测率可达92%以上。

To solve the problem of low precision in detecting multi-target abnormal behaviors such as fighting and robbery,in this paper,a novel human abnormal behavior detection algorithm is proposed based on a combination of the weighted reconstruction trajectory and histogram entropy.First,the background subtraction method is combined with aspect ratio to extract pedestrian targets,and then Kalman filters and histogram of oriented gradient(HOG)features are integrated into a spatio-temporal context algorithm to realize a robust tracking of completely occluded pedestrians in a short time,and finally the tracked trajectories are trained;a normal behavior directory is constructed for the motion trajectories defined as normal,and the trajectories to be tested are sparsely reconstructed.An effective detection of abnormal behavior is realized by combining the weighted minimum reconstruction residuals and amplitude direction histogram entropy.A comparative experiment shows that the algorithm can effectively detect abnormal behaviors such as fighting and robbery,and the detection rate can exceed 92%under static background without occlusions.