Research And Application Of Surveilance Video Abnormal Event Detection Based On Deep Neural Network

With the advancement and construction of projects such as smart cities,safe cities,and safe mines,surveillance cameras,as the main collection tool for visual signals,have been placed in large numbers in all corners of cities and industrial areas.These cameras generate massive surveillance videos related to security protection at all times,which urgently need to be analyzed and processed in a timely manner.As an automatic visual feature extraction technology,deep convolutional neural network not only greatly improves the representation ability of the visual model,but also brings a new methodology in the design of visual algorithms.Considering that convolution operators are generally only sensitive to the features of two-dimensional images,and video is a typical three-dimensional spatio-temporal data structure with time series,it is difficult for traditional convolutional networks to fully mine video frames in the time dimension.And get effective video feature representation.Based on this,this paper proposes a Multi-semantic Longrange Dependencies Capturing(MLDC)method to enhance the timing modeling capabilities of convolutional neural networks.This method first breaks each frame of the video into multiple units with strong semantic information,and then uses a sub-network based on the attention mechanism to model the cross-sequence dependency between semantic units.Finally,this paper designs and implements a real-time detection system for surveillance video abnormal events,and applies the system to the detection task of roadway water inrush accidents.Through sufficient experimental analysis and real water inrush experiments,the effectiveness of the proposed algorithm and system is verified.The work of this paper is summarized as follows:(1)Through the analysis and comparison of video feature representation algorithms based on dependency capture,it is proposed to introduce more semantic information when modeling the dependency relationship between video frames,and based on this,a multi-semantic long-range dependency capture algorithm is designed.By embedding the algorithm iteratively into the commonly used deep convolutional network,the network's ability to identify video moving subjects is effectively improved,and the quality of video feature representation is enhanced.An Early Dependencies Transferring(EDT)technology is proposed to speed up the inference speed of the MLDC algorithm.(2)Perform detailed digestion experiments,comparative experiments and visual analysis on the proposed method on large-scale video understanding data sets such as Kinetics-400 and Something-something v1.The experimental results show that the MLDC algorithm not only achieves the best recognition and classification accuracy,but the MLDC algorithm accelerated by the EDT technology also meets the real-time requirements.The results of the visual analysis further show that the proposed method effectively improves the network's ability to recognize moving subjects and time series modeling.(3)Using Web technology and sensor technology,a universal real-time detection system for abnormal surveillance video events is built,which is mainly composed of a sensor network module,a system Web module,and a deep learning algorithm service module.The system is applied to the task of water inrush detection in roadway.Through the simulation of real water inrush experiment,it is proved that the system has high performance in the task of water inrush detection.