Research On Indoor Human Abnormal Behavior Recognition Method Based On Space-Time Information

In recent years,with the increasing demand for public area security,based on the intelligent analysis of video surveillance equipment application scenarios continue to expand,indoor security scenarios of human abnormal behavior recognition is one of the important application areas,such as bank vaults,archives and other confidential places,indoor abnormal behavior put forward a more stringent control needs.With the continuous improvement of hardware performance,some of the video surveillance equipment has been equipped with intelligent analysis module,which can be achieved based on video data analysis of personnel posture,simple behavior recognition and other functions.Existing video surveillance systems have been widely used on some occasions with low behavior supervision needs,but in indoor security places with high behavior supervision needs,it is difficult to achieve effective identification for specific abnormal behavior,and identification accuracy is challenging to guarantee.Therefore,it is important to study a robust indoor human abnormal behavior recognition system.This topic takes the fall and brawl behaviors in the bank vault scene as the recognition object.Based on the human key point detection to achieve the extraction of skeleton information of people,it focuses on the single person abnormal behavior recognition problem associated with fall detection and the two-person interaction abnormal behavior recognition problem associated with brawl detection.The details are as follows.(1)Overall scheme design of indoor human abnormal behavior recognition system.Combined with the analysis of detection scene characteristics and technical index requirements,complete the selection of equipment such as cameras,servers,recorders and hardware system design.Based on the analysis of the recognition requirements of abnormal behavior,the software system operation process and the process design of the behavior recognition algorithm based on video analysis are completed.In order to improve the recognition accuracy,the camera scheduling method based on the optimal detection viewpoint determination is proposed.(2)Human key point detection algorithm based on improved Open Pose.For the problem that there are more interference prospects in complex scenes,a foreground person extraction method based on human key point detection is proposed.In order to achieve the global optimal matching of key points in multi-person scenes,the improved Hungarian algorithm does bipartite graph matching of the detected key points and solves the optimal matching by augmented path traversal;meanwhile,a Gaussian kernel-based At the same time,a Gaussian kernel-based key point labeling strategy is proposed to make the description of the distribution of joint positions in the labeled graph closer to the actual situation;in order to guarantee the integrity of the coordinate vectors in the skeletal sequence,the missing value compensation is achieved by nearest neighbor weighted regression.(3)A single person abnormal behavior recognition method based on channel topological transform CNN(Convolutional Neural Network).The fall behavior is a typical single person abnormal behavior,and the recognition method based on the improved CNN is proposed for the recognition error problem caused by the existence of the interference of similar behaviors of motion patterns.In order to amplify the limb motion amplitude difference of similar behaviors,the pose feature description of the person is refined through motion correlation feature modeling;a channel topological transform CNN with velocity feature extraction capability is designed to further achieve effective differentiation of similar behaviors based on the limb motion frequency difference.Meanwhile,in order to reduce the time overhead generated by processing redundant frames,a key frame extraction strategy based on the analysis of nodal motion characteristics is proposed.(4)A method for identifying abnormal behavior of two-person interaction by fusing spatiotemporal graph sequence features.A typical two-player interaction behavior is brawling,and a recognition method is proposed to fuse spatio-temporal map sequence features to address the problem of low recognition accuracy due to mutual occlusion of people.In order to enrich the description of the connection state between joints in the case of occlusion,the intra-frame spatial map and inter-frame temporal map connection design is carried out based on the skeletal sequence information,and the spatio-temporal map feature descriptor of brawling behavior is constructed;in order to better combine the temporal sequence information,the spatio-temporal map sequence features are further extracted by improving the long and short term memory model: to address the problem of insufficient decision complementarity of adjacent temporal sequence nodes of the model,the threshold value complementary sharing The internal loop unit of the model is optimized based on the threshold value complementary sharing strategy;in order to constrain the state distribution of the model parameters during the training process and guarantee the linear circulation of gradients between layers,layer normalization and linear scaling residual connection branch design are introduced in the model.Above all,this paper completes the design of an indoor human abnormal behavior recognition system,and investigates human key point detection,fall behavior and brawl behavior recognition methods.On the basis of this research,the development of the system software is further completed.The software mainly includes functional modules such as user login,automatic operation,offline testing and data management.To verify the effectiveness of the recognition algorithm,based on the test analysis of the collected behavioral data samples,the accuracy rates of fall and brawl behavior recognition are 95.2% and 93.4%,respectively,with the over detection rate less than 5% and the omission rate less than 1%,meeting the performance index of abnormal behavior recognition.To verify the detection efficiency of the system,a quantitative analysis of the detection elapsed time was performed,the average image processing time per frame is less than 110 ms,which meets the detection speed requirement of more than 5 frames per second.Moreover,the system software runs robustly and can meet the usage requirements.