Design And Implementation Of Abnormal Behavior Recognition System Based On Deep Learnin

Abnormal action recognition is an important branch in the field of action recognition,which recognizes specific human action and then judges whether it is abnormal.Timely identification of abnormal human action is of great significance in the fields of smart nursing,intelligent monitoring,and sports assistance.Therefore,designing an abnormal action recognition system has broad application prospects and important practical application value.At present,deep learning has become the mainstream of action recognition methods due to its powerful feature extraction capabilities and fast data processing capabilities,but the actual implementation of action recognition algorithms still has problems such as long sequence modeling difficulties,complex models,and poor robustness.Compared with data modalities such as RGB,the human 3D skeleton contains richer semantic information in less data and has good robustness to environmental lighting changes.Based on the above background,this paper proposed two action recognition networks based on human 3D skeleton and self-attention mechanism,and designed and implemented an abnormal action recognition system.(1)Aimed at the characteristics of compact structure and long sequence of human skeleton data,a two-stream spatial-temporal self-attention network based on selfattention mechanism was proposed to overcome the problem of attention loss in long sequence modeling of RNN and LSTM.The network space flow extracts the spatial structure features of human joints,and the time flow extracts the timing characteristics of the joints action process;used self-attention to model the temporal-spatial change relationship of joints,used multi-head attention to focus on different features from different subspaces,and used Sum fusion strategy fusion the prediction scores of the two streams;the results of the joint motion and joint feature input model were fused.Verified by the NTU RGD+D dataset and SHREC 2017 dataset,the action recognition effect was good,The results of the experiment proved the effectiveness of the proposed network.(2)Aimed at solving the problem that the dual-stream spatial-temporal self-attention network model has good recognition effect but high computational cost,and the practical implementation of the multi-feature fusion scheme has certain difficulties,a lightweight model of the single-stream scheme was proposed,based on the short-sequence enhanced network model of Transformer.This model improved the weak modeling ability of single-stream networks in the time dimension by adding joint motion modules and shortsequence enhancement modules to spatial streams.The main body of the model has 2modules,the joint movement module,which calculates the difference between each frame and adjacent frames.Assign different weights to all the differences and the features of the joint point itself as the input of the attention module;the short-time sequence enhancement module aggregates the features of the adjacent frames of the joints through a sliding window.The model was verified on the NTU RGD+D bone dataset,and a good action recognition effect was achieved.(3)The design and implementation of the abnormal action recognition system have been completed.The system includes four functions: real-time recognition of abnormal behavior,offline recognition,data set expansion,and video recording.An abnormal action dataset has been established and retrained the Transformer-based short-sequence enhanced network on the builted dataset,and applyed the best model to the abnormal action recognition system.Then analyzed the system requirements,designed the software program and system UI of the system,and realized the related functions of the abnormal behavior identification system.Finally,the functions of each module of the system are tested through experiments,and the accuracy,generalization,robustness and real-time performance of the system were verified.