Research On Violence Recognition Algorithm In Video

Violent behavior recognition is an emerging technology based on traditional intelligent video surveillance.It is an interdisciplinary filed for computer vision,pattern recognition and machine learning,and its purpose is to accurately detect and identify the violent actions in surveillance videos.The research has a value to be reckoned with improving the application of video surveillance systems in the security field.Inheriting traditional violent behavior recognition methods to extracting apparent characteristics and motion features,this thesis combines them with deep learning frameworks based on convolutional neural networks.Research work is focused on key frame extraction,multiple feature extraction and fusion,as well as the acceleration of violence detection algorithms in videos.1)For the redundancy of apparent features and the sparsity of temporal distribution of violent behavior characteristics in videos,this thesis proposes a violent behavior evaluation algorithm based on local differential brightness to roughly sift the video keyframes,by which the redundancy of frames is reduced while the effectiveness of the input video information is enhanced.Meanwhile,gaussian distribution is applied to the selection of training frames,which increases training sample combinations and avoids the problem choosing only those samples with higher separability for training.The experimental results indicate that this approach can effectively screen the features of the input videos,which improves the recognition accuracy of the system and preliminarily solves the mismatch problem.2)For the problem that the characteristics of video keyframes cannot fully represent the violent behavior features,this thesis combines the key frame selection work and proposes a PVFN network to integrate video frame features.Residual network is used as the infrastructure to enhance the violent features and weaken the non-violent ones by the video frames merging in the spatial feature dimension.In addition,a two-stage video feature extraction structure for the residual network partitioning is introduced,which improves the single-channel expression capability with hardly any network parameters increasing,and thanks to the jumping connection,the fusion of frame features is augmented.The experimental results demonstrate that the network substantially boosts the accuracy of violence identification.3)In view of the non-repetitive character of violent behavior and the poor real-time performance of the violent behavior recognition system,this thesis analyzes the violent behavior representation the real-time calculation ability of three different motion characteristics:optical flow,differential image and ALMD The experiments exhibit that with the inputing of other lightweight motion features,the real-time performance of the system can be improved while maintaining close proximity to the optical flow indices.In order to better evaluate the effectiveness of the algorithm,this study designs a largescale violence identification dataset to compensate for the lack of standard datasets under monitoring conditions currently.Comparison of proposed brutal force recognition algorithms with other ones show that the former achieves higher accuracy and stability in both collected dataset and multiple benchmarks.Moreover,a practical video surveillance violence identification software and hardware system is constructed to demonstrate the real-time capability and robustness of the algorithm.The comprehensive experimental results indicate that the introduced algorithm is of great research significance and application value.Part of the code and test results for this thesis can be found here:https://github.com/DavidZh666/Violence_Detection.