Research And Application Of Human Action Detection In Real Scenes

The rapid development of deep learning technology has spawned a series of practical application scenes and directions such as computer vision,natural language processing,and reinforcement learning.At the same time,the development of DL in the field of secu-rity monitoring has also ushered in changes in technical means.However,most of the DL algorithms currently applied in the field of security monitoring only stay in the laboratory stage.Although the current algorithms can achieve a good accuracy for commonly used data sets,the accuracy and real-time performance of the algorithm are in real scenes.Can not meet the requirements of practical applications,so an intelligent human behavior de-tection system is urgently needed to solve the current human behavior detection tasks in real scenes.Based on human behavior detection tasks in real scenes,this paper uses DL technology to design a set of detection algorithms and recognition algorithms to analyze behaviors.At the same time,a set of distributed behavior detection platform is developed to realize a set of video Analyze the end-to-end system that detects alarms.The specific work is as follows:1.Designed a human behavior recognition algorithm named VSTNet network.This network abandons the traditional CNN feature extraction method,draws on the Trans-former architecture in the NLP field and the residual network solution for gradient dis-appearance,and applies the Seq2 Seq solution in the text processing field.In the field of image and video processing,a video image feature extraction network based on the Attention mechanism is designed.At the same time,considering the unique timing fea-ture extraction problem in the field of video behavior analysis,a timing feature extraction module is designed,named PSM,which abandons the traditional 3D convolution or opti-cal flow method is an inefficient time-series feature extraction method,manually shifting the features to merge the features between different frames to perform temporal context modeling,so that the timing modeling can be solved efficiently.2.Designed a human behavior detection algorithm named MTResnet network.The network uses Res Net50 as the backbone network.At the same time,the PSM module is embedded in the residual branch of the residual model bottleneck structure to achieve the extraction of timing features.For small targets in real scenes For object detection,this paper uses the feature pyramid model to restore high-resolution features through upsam-pling.Finally,this model replaces the region generation network of the traditional target detection network through the Transformer's codec architecture,avoiding the generation of anchor frames.On the basis of concise processing,results comparable to the current excellent models have been achieved.3.Aiming at the designed human behavior analysis algorithm,a distributed back-end monitoring system with C/S architecture is designed.The system can dynamically call the back-end server to detect and achieve load balancing according to requirements,which solves the shortage of GPU resources in the deep learning system and the problem of uneven load on service nodes,proposed a feasible solution for the construction of large-scale security systems on the market.