Video Crowd Behavior Detection Based On Spatial-temporal Invariant Features

With the urbanization of human society and growth of human population,it becomes more and more important of intelligent management in the area of city public security,anti-terrorism.Many HD video monitoring systems are deployed here and there.People pay more and more attentions on how to make use of these massive information effectively and detection these public security incidents in advance.Combined with reality demands,the paper focuses on the detection and recognition of crowd violence behavior in surveillance videos.After researching spatial-temporal features and dictionary learning method,a video crowd behavior detection algorithm based on spatial-temporal feature is proposed.Based on the algorithm,an violence behavior detection system is also be developed.The main work of this paper can be concluded as follow:1)We research and evaluate different violent video spatial-temporal features.The paper compare different spatial-temporal features, such as MoSIFT,ViF and SC-ISA(Stacked Convolutional ISA feature);discuss the difference between hand-design features and unsupervised learning features.2)A new dictionary-learning algorithm based on Stacked Fisher Vector Encoding is proposed.The algorithm has the advantage of high dictionary-learning speed and discriminative mid-level representation.3)A new violent behavior detection algorithm based on SC-ISA and SFV is proposed.Evaluation experiments are tested in the public crowd violence dataset,the accuracy and detection speed both outperformance other state-of-the-art algorithms.4)A violent crowd behavior recognition and detection platform base on deep learning ideas is developed.The system has a good performance on the detection of violence behavior in pulic places.