Research On Two Stream CNN Based Abnormal Behavior Classification

With the close attention to public security,monitoring equipment is popular in public contexts.In consequence,the number of monitor videos is increasing with a high speed.Though,observing the monitoring videos by handcraft is inefficiency and costly.Therefore,it’s an important research topic in computer vision to classify the human behavior and distinguish whether the behavior is abnormal.In this paper,we focus on the research of classifying abnormal behavior with two streams CNN which uses optical flow diagrams and RGB video frames as inputs.Compare to the application of specific artificial feature extraction,deep learning have a stronger learning ability,but the training process of the CNN requires lots of training data,and easily lead to over fitting.In order to solve these problems,we use different sample translation methods and modify the net structure properly to gain a better result in the abnormal behavior classification task.This paper integrates different datasets to construct a new dataset which covers the common actions in life for abnormal behavior classifying.This paper selected the feature of optical flow to represent dynamic information in video frames.On account of the low efficiency of traditional optical flow calculation,we proposes adaptive optical flow computing method to represent the motion of human behavior efficiently and accurately,and transform optical flow vector to color space image.Because of the complexity of the task and the shortage of video dataset,this paper also uses different transform methods such as cropping and flipping as information entropy methods.In the network model,this paper adjusts the network structure to improve the accuracy.Finally,this paper compares the classification results of the improved network model and the original model and artificial action features.We also define the abnormal behavior under different context to finish the task of abnormal behavior classification.