Research And Implementation Of Abnormal Action Recogniztion And Detective Based On Deep Learning

Nowadays,video surveillance systems are widely used,but for traditional video surveillance,their self-recognition ability is weak.It is difficult to truly meet the needs of the current society because of their excessive dependence on human processing,so a more intelligent video system is needed.The monitoring system monitors abnormal behaviors in real time to solve the problem of weak self-recognition ability of the current monitoring system.Therefore,based on the premise of using in real scenes,cascad convolutional neural networks are proposed in this thesis to detect and locate abnormal behaviors for real surveillance videos.Meanwhile,a set of intelligent monitoring platform was developed,and the abnormal behavior algorithm was applied to the monitoring platform.In this thesis,a deep learning algorithm based on convolution neural network for abnormal behavior recognition and location is studied and implemented,the specific work is as follows:1.A cascade network for abnormal behaviors detection is designed.The first phase is the preliminary behavior screening phase.This phase allows high fault tolerance.It is named as FANet(Fighting Attention Net)network.The network uses the improved residual network as the basic feature extraction network.Multi-scale feature fusion is achieved by connecting feature layers of different scales;more asymmetric convolutions are introduced into the feature network to reduce the amount of network calculations.RPN is used to optimize and classify and locate the behavior,and ROI Align is used to make the feature map output at a fixed size.When the behavior score reaches the threshold,continuous video frames are intercepted and transferred to the second stage network based on the starting point of the abnormality.The second stage network is named MP-R3D((MultiPath-Res 3D Convolutional Net)).This network uses a multi-fiber network based on 2D convolution in the shallow layer and 3D convolution in the deep layer to reduce the complex spatiotemporal fusion and huge 3D convolution during training.Memory consumption without reducing behavior recognition rate.2.The two levels of the cascade network are trained and fused independently,so that they can to recognize behaviors.Modified the input method of the network,which increased the GPU utilization by 44.2%.Instead of directly obtaining the RTSP videostream,a video decoder is used to enable the detection network to perform multiple concurrent detections.A data set for abnormal behavior was proposed.Unlike most short-clip public data sets,this data set includes an image-based dataset set with coordinate information and a clipped video set.The rich and diverse data set is closer to the real scene.3.Combined with the designed abnormal behavior recognition algorithm,a video surveillance system is designed.The monitoring system is based on the C / S architecture design.When the network video stream is connected to the platform,it can monitor the abnormal behavior in real time.When an abnormality occurs,it will send an alarm to the end user as soon as possible,and keep relevant video clips and alarm logs to facilitate Time to go back.