Fighting Behavior Detection In Surveillance Video By Two-stream Convolutional Networks

In order to protect human safety,surveillance has spread all over the streets.Usually,our monitoring can only record the pictures and need to monitor the occurrence of abnormal behaviors.This method is not only inefficient but also costly.Scholars have always hoped that abnormal behaviors can be automatically monitored.Therefore,video-based abnormal human behavior recognition has always been an important and challenging research direction in computer vision.In the past,research on video has studied the identification of abnormal behaviors such as shooting,brandishing,slashing,etc.However,due to the randomness and unpredictability of fighting behavior,there is no good result in the research on the identification of fighting behavior.In addition,with the development of deep learning,the convolutional neural network has become the most advanced image classification model.Therefore,based on this,the paper identifies the fighting behavior in video data and obtains good results.The main work of this paper is as follows:(1)Aiming at the problem of fighting behavior identification with the deep learning method,this paper proposes a fighting behavior identification method based on the Two-Stream convolutional neural network,which is mainly based on the Two-Stream convolutional neural network model.Mainly because of the Two-Stream convolutional neural network makes full use of the temporal and spatial components of the video to provide more motion information for identifying fighting behavior.(2)This paper proposes an improved Two-Stream convolutional neural network model for the identification of fighting behavior in surveillance video for the problem of the large amount of parameters and low operating efficiency of the original Two-Stream convolution model.One of the improved ideas is to reduce the number of convolution layers.On the other hand,the two convolutional layers of the Two-Stream model are merged before entering the last layer of fully connected layers.Thus,A more simple and efficient Two-Stream convolutional neural network for identifying fighting behavior is obtained.The average recognition rate can reach 89%.(3)In this paper,we use the Boltzmann entropy to detect abnormal frames in the video for the need to make a large amount of fighting video frame data,so as to pre-classify the fighting and non-fighting frames of the collected fighting video data.Because of the probability of anomalous events is small,we first use the Boltzmann entropy in the normal state to establish a Gaussian model,and then use this Gaussian model to calculate the probability of the test sample to determine the occurrence of abnormal events.(4)In order to improve the recognition rate of fighting behavior in the method of this paper,the optical flow frames are processed by superimposing 5 frames and 10 frames respectively,and then training experiments are carried out.Because of the superposition of the optical flow enhances the human movement trend,it can be proved that training on multi-frame stacked optical flows can achieve a better fight recognition effect.Finally,this paper compares the traditional method HOG+SVM and the Two-Stream network extensions VGG16,ResNet and GoogleNet.It validates the feasibility and effectiveness of the deep learning method proposed in this paper to identify fighting behaviors.