智能监控中目标检测与异常行为识别算法研究

发布时间：2018-04-15 11:21

本文选题：智能监控 + 目标检测　；参考：《西安建筑科技大学》2017年硕士论文

【摘要】：为了解决传统视频监控工作效率低下、耗费人力物力等问题,本文通过对视频数据流的分析,自动实现场景中的目标检测,并在此基础上进行目标行为分析、识别和理解,实现在监控场景下发生异常行为的智能监控和处理。具体的工作如下:1)根据几种经典的运动目标检测算法构建一个框架,该框架在考虑性能和设计时间成本的情况下,随着场景的变化,自适应地选择最优算法参数组合对运动目标进行检测,并且运行时间能够根据实时输入进行自适应调节。有效解决不同场景下运动目标检测效果差异明显的问题,从而增强检测算法的通用性。2)提出一种基于模糊迭代自组织数据分析聚类结合直方图熵值算法对异常行为进行检测。该算法通过模糊迭代自组织数据分析聚类方法获取视频关键帧并分类,从而减少计算复杂度。对三种异常行为进行建模,根据分类结果和建立的异常行为模型,采用直方图熵值作为总特征对异常行为进行检测,并与方向梯度直方图结合支持向量机的异常行为检测算法进行比较分析。3)在不同场景下,分别对三种异常行为的运动特征进行分析,获取异常行为特征参数。设定三种异常行为判定准则,将异常行为检测模型所定义的规则存储到数据库中,建立异常行为规则库,给出异常行为规则匹配流程,分析检测出的异常行为是否与规则匹配。实验表明,所建立的异常行为规则库能够识别出具体异常行为。综上所述,在智能视频监控系统中构造的自适应检测算法框架能够解决单个检测算法通用性不强的问题,文中提出的模糊迭代自组织数据分析聚类结合直方图熵值的异常行为检测算法相对于方向梯度直方图结合支持向量机算法检测准确率更高,建立的异常行为规则库能够匹配出异常行为。该异常行为识别算法可以有效地提高监控的效率,对维护社会治安的稳定具有实际应用价值。
[Abstract]:In order to solve the problems of low efficiency of traditional video surveillance and the cost of manpower and material resources, this paper analyzes the video data flow, realizes the target detection in the scene automatically, and on this basis carries on the target behavior analysis, the recognition and the understanding.Realize the intelligent monitoring and processing of abnormal behavior in the monitoring scene.The specific work is as follows: 1) based on several classical moving target detection algorithms, a framework is constructed, which takes into account the performance and design time cost, as the scene changes.The optimal algorithm parameters are selected adaptively to detect the moving target, and the running time can be adjusted adaptively according to the real-time input.In order to enhance the generality of the detection algorithm, an algorithm based on fuzzy iterative self-organizing data clustering and histogram entropy algorithm is proposed to detect abnormal behavior.The algorithm uses fuzzy iterative self-organizing data analysis clustering method to obtain video key frames and classify them, thus reducing the computational complexity.According to the classification results and the established abnormal behavior model, the histogram entropy is used as the general feature to detect the abnormal behavior.And compared with the direction gradient histogram and support vector machine anomaly detection algorithm. 3) in different scenarios, the motion characteristics of three abnormal behaviors are analyzed, and the abnormal behavior feature parameters are obtained.The rules defined by the anomaly detection model are stored in the database, the abnormal behavior rule base is established, the matching flow of the abnormal behavior rule is given, and whether the detected abnormal behavior matches with the rule is analyzed.The experimental results show that the rule base of abnormal behavior can recognize the specific abnormal behavior.To sum up, the adaptive detection algorithm framework constructed in the intelligent video surveillance system can solve the problem that the single detection algorithm is not universal enough.The proposed fuzzy iterative self-organizing data analysis clustering combined with histogram entropy detection algorithm is more accurate than the directional gradient histogram combined with support vector machine algorithm.The rule base of abnormal behavior can match the abnormal behavior.The algorithm can effectively improve the efficiency of monitoring and has practical application value in maintaining the stability of public order.
【学位授予单位】：西安建筑科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【参考文献】