摘要
针对传统中小群体异常行为检测方法实时性与场景适应性不能达到平衡的问题,从机器学习角度出发,提出一种通过群体密度特征和运动特征来进行群体异常行为检测的方法。利用快速群体密度估计方法提取群体密度特征,采用局部稠密光流法提取群体运动特征,结合密度特征,利用随机森林算法完成对群体异常行为的识别。在PETS和UMN公共数据集上进行实验,与传统群体异常行为检测方法进行对比,实验结果表明,该方法能够有效检测出中小群体异常行为,识别其类型,识别率能达到97%,实时性好,鲁棒性强。
In the traditional methods of detecting the abnormal behavior of the small and medium crowd,the detected type of abnormal behavior and real-time cannot achieve balance.Starting from the perspective of machine learning,an abnormal crowd behavior recognition method based on the crowd feature of density and motion was proposed.The feature of crowd density was obtained using rapid crowd density estimation methods.The feature of motion was obtained based on local dense optical flow method.The random forests were used to complete the recognition of abnormal crowd behavior combined with the feature of crowd density.Experimental results on the PETS and UMN datasets show that the method can detect and recognize different crowd abnormal behaviors with 97% accuracy,which is higher than traditional methods,and the proposed method is real-time and robust.
出处
《计算机工程与设计》
北大核心
2016年第9期2507-2514,共8页
Computer Engineering and Design
基金
教育部"春晖计划"基金项目(Z2012029)
四川省信号与信息处理重点实验室开放基金项目(szjj2012-015)
西华大学研究生创新基金项目(ycjj2015098)
关键词
异常行为
机器学习
群体密度
光流
随机森林
abnormal behavior
machine learning
crowd density
optical flow
random forests