Aim at the performance of crowd behavior analysis had a common feature extraction at the present stage, the results of crowd behavior analysis can't reach the requirement of surveillance analysis, this article proposes a motion feature based on the group level of crowd, and the motion feature extracted from collectiveness, stability and conflict.Then, importing the three groups of motion features in the convolution neural network for training, after training the algorithm getting the behavior labels for crowd.At last, some contrast experiments were comparing on Violence dataset.Experimental results show that, the proposed group-level motion features will give the most basic features for crowd, it will build discriminative features for recognition, and the behavior analysis result of Violence show that group-level motion features can extend into independent-scene, for any scene it can obtain robust basic motion feature.And the convolution neural network for training and recognition will improve the accuracy for crowd behavior analysis, compared with convention algorithms, the proposed algorithm will perform in independent scene, and improved the average accuracy of 13%, this research will do contribution for surveillance scene crowd behavior analysis.%针对现阶段人群行为分析的特征提取效果不佳,人群行为分析结果达不到视频分析的要求.提出一种基于人群群组级别的动量特征,分别表示人群的集体性、稳定性和冲突性,然后将三组人群群组动量特征输入至卷积神经网络进行训练,最后在Violence数据集上进行人群行为分析实验.实验结果表明,提出的群组动量特征能够在群组级别表达出人群的基本特性,这些特性在人群行为分析中能够建立可识别较高的特征,在Violence数据集上的测试结果显示.提出的群组动量特征能够扩展到独立场景,对于任何场景的人群行为分析都能够获得鲁棒的基础动量特征,而采用卷积神经网络进行的训练和分类,能够提升人群行为分析的精确度.与传统特征及分类方法相比,在各种不同的独立场景中,将标注结果精度提升了13%左右,在视频场景人群行为分析中具有较强的实践意义.
展开▼
