基于机器学习的视频序列中自动人数统计研究

发布时间：2018-10-09 21:27

【摘要】：智能视频监控技术日益成为公共安全管理的重要技术手段,自动人数统计是智能视频监控的重要内容之一,该技术的研究对于实现平安社会具有重要意义。由于视频中含有大量信息,现有方法在处理时未能同时兼顾准确性与实时性。针对这一问题,论文主要的研究内容如下:针对场景中目标运动速度较快时跟踪效果较差的问题,提出了一种改进的多目标跟踪方法。该方法在使用背景消减法完成对前景目标的提取后,对目标进行Kalman预测,以该结果作为Mean-shift搜索的初始位置展开迭代,并将其结果作为Kalman修正阶段的观测值,从而完成对Kalman滤波器的更新;另外,引入遮挡因子对目标的遮挡情况加以判别,从而实现对遮挡的自适应处理。实验结果表明,该方法可以有效减少跟丢的发生,对场景中运动速度较快的目标具有较好的鲁棒性,并且在不同帧中均能较好实现对同一目标的匹配。针对提取的人体特征维数较高而造成的运算过程复杂等问题,提出一种基于粗糙集的多特征约简方法。提取一系列具有代表性的目标特征,然后通过粗糙集中知识约简的方法对特征进行约简。实验结果表明,该方法能有效减少识别时间,满足视频处理中对实时性的要求。另外,对于提取的行人特征,研究了基于机器学习的人数统计方法,使用自适应动量因子对BP算法进行了优化,该方法引入自适应动量因子对各层之间的权值进行更新,从而完成误差的反向传播。实验结果表明,该方法能够有效改善由于常动量因子取值不当造成的BP算法不稳定,并具有较低的时间复杂度。论文对视频序列中自动人数统计的研究方法能够充分利用监控视频中的信息数据,实现对人体目标的自动检测及跟踪,并掌握精确的出入人数,从而杜绝高密度人群场所易出现的安全隐患,对公共场所的管理以及人群灾祸预防具有重要意义。
[Abstract]:Intelligent video surveillance technology has increasingly become an important technical means of public security management. Automatic population statistics is one of the important contents of intelligent video surveillance. The research of this technology is of great significance for the realization of a safe society. Due to the large amount of information in the video, the existing methods fail to take both accuracy and real-time into account. Aiming at this problem, the main contents of this paper are as follows: aiming at the problem of poor tracking effect when the target moves faster, an improved multi-target tracking method is proposed. After the foreground target is extracted by background subtractive method, the target is predicted by Kalman, the result is used as the initial position of Mean-shift search, and the result is used as the observation value of Kalman correction stage. In addition, the occlusion factor is introduced to judge the occlusion of the target, and the adaptive processing of occlusion is realized. The experimental results show that the proposed method can effectively reduce the occurrence of loss and is robust to the targets moving faster in the scene, and the matching of the same target can be achieved in different frames. In view of the complexity of the operation process caused by the high dimension of human feature extracted, a multi-feature reduction method based on rough set is proposed. A series of representative target features are extracted, and then the features are reduced by rough set knowledge reduction. Experimental results show that the method can effectively reduce the recognition time and meet the real-time requirements in video processing. In addition, for the extracted pedestrian features, the population statistics method based on machine learning is studied, and the adaptive momentum factor is used to optimize the BP algorithm. The adaptive momentum factor is introduced to update the weights between layers. In order to complete the reverse propagation of the error. Experimental results show that the proposed method can effectively improve the instability of the BP algorithm caused by the improper value of the constant momentum factor and has a lower time complexity. The research method of automatic number statistics in video sequence can make full use of the information data in the surveillance video, realize the automatic detection and tracking of human body target, and grasp the accurate number of people entering and leaving. Therefore, it is of great significance for the management of public places and the prevention of crowd disasters to put an end to the safety hidden dangers which are easy to appear in high density crowd places.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181;TN948.6

【参考文献】