结合在线随机张量分解和全变分的运动目标检测

Moving object detection combines OSTD and TV

移动摄像机拍摄的视频由于背景帧不固定以及数据量大的问题,有效检测出其中的运动目标仍是一个挑战。提出了结合在线随机张量分解(OSTD)和全变分(TV)正则化的运动目标检测方法。首先,将视频序列可视为三阶张量,采用在线优化方法实时处理。结合小样本批处理初始化缩小基矩阵大小,在前景部分利用TV正则化对目标的空间性进行约束。最后,不断更新基矩阵和稀疏部分系数,直到达到设定的迭代次数或所有样本计算已经完成。实验结果表明:该方法能有效应对背景帧不固定、实时处理等问题,准确地检测出运动目标。与其他同类方法相比,不仅检测精度更高,而且保持了较快的运行速度。

Due to the problem that the background frame is not fixed and the amount of data is large,it is still a challenge to effectively detect moving objects captured by moving camera video.A method of moving object detection based on online stochastic tensor decomposition(OSTD) and total variation(TV) regularization is proposed,Firstly,the video sequence is regarded as a third-order tensor,and the online optimization method is used to process the video sequence in real-time.Combined with the small sample batch processing,the size of the base matrix is reduced.In the foreground part,the space of the target is constrained by TV regularization.Finally,the base matrix and sparse part coefficients are updated continuously,until the set number of iterations is reached or all samples are calculated.Experimental results show that the method can effectively deal with the problems of background frame instability,real-time processing,etc.,and effectively detect moving objects.Compared with other similar methods,it not only has higher detection presision,but also maintains a faster running speed.