基于蒙特卡洛树搜索的视频异常场景监测方法

Method of video abnormal scene monitoring based on Monte Carlo tree search

通过视频监控可以更快速地发现异常场景,并尽快制止违法暴力行动。为保证监测精度,文中设计一种基于蒙特卡洛树搜索的视频异常场景监测方法。首先对行人轨迹特征进行提取,计算目标预测位置与实际位置的空间距离,判定二者的相对位置,建立三级异常图像;再基于蒙特卡洛树搜索算法设计行人行为判别方法,构建不确定判别网络,标记残差函数,对数据进行归一化处理,同时获得激活函数以及函数的输出值;最后,设计视频异常场景监测算法,基于协方差矩阵将多个判别结果汇总成一个整体,并以此得到监控视频内异常场景的监测结果。该方法能够通过目标提取得到视频内的异常目标。目标识别精度的测试结果表明,所提方法在简单场景与复杂场景内的AUC值分别为0.952和0.886,说明其监测精度较高,在简单场景与复杂场景下均可正常使用。

By means of video monitoring,abnormal scenes can be found more quickly and illegal violence can be stopped as soon as possible. A video abnormal scene monitoring method based on Monte Carlo tree search is designed to ensure the monitoring accuracy. The pedestrian track features are extracted,the spatial distance between the predicted position and the actual position of the target is calculated to determine the relative position of them,and a three-level abnormal image is established. And then the pedestrian behavior discrimination method is designed on the basis of the Monte Carlo tree search algorithm,the uncertain discrimination network is constructed to mark the residual function and conduct the normalization processing of data,so as to obtain the activation function and the output value of the function. A video abnormal scene monitoring algorithm is designed,and on the basis of the covariance matrix,multiple discrimination results are summarized into a whole,thus obtaining the monitoring results of abnormal scenes in the monitoring video. This method can be used to obtain the abnormal target in the video by means of object extraction. The results obtained in target recognition accuracy testing show that the AUC values of the proposed method in simple scenes and complex scenes are 0.952 and 0.886 respectively,which indicate that its monitoring accuracy is high and it can be used normally in simple scenes and complex scenes.