Action Recognition With Convolution Neural Networks And Parts

With the rapid development of economy and society,there are more and more surveillance cameras in our life.The police departments use the monitoring record to solve cases,the traffic departments use it to capture illegal vehicles,while supermarkets use it to monitor the safety of goods and so on.However,these monitoring records are mostly used for later review,which is relatively inefficient.In order to solve the problem of lower using efficiency of video surveillance,we can integrate behavior recognition technique into video surveillance to warn some illegal guys and bad actions in real time,which will greatly improve the use efficiency of video surveillance.And,this will outstrip the efficient of manual screening.One of the core research fields is behavior recognition in the field of machine vision currently.The main task of action recognition is to identify or estimate what actions people execute in videos,and even predict the actions about to happen.In real life,the recognition effect of these videos is still unsatisfactory even when the model with good experimental results is used to identify the behavior of these video,which is caused by illumination,occlusion,weather,the shooting quality of the camera probe itself and so on.Although the traditional method has made considerable progress,it is not feasible to apply it to the actual situation due to the excessive consumption of computational resources.In recent years,with the emergence and development of Deep Learning,those problems has been largely alleviated.Faster,more efficient and more accurate recognition of human behavior has become a goal that is going to realize.In this paper,action recognition technology is studied.The main work is as follows:(1)In this paper,a method of behavior recognition based on convolution neural network and body parts is proposed.The classification of behavior is synthetically determined by the ability of convolution neural network which could automatically extract features and the detailed information earried by parts.Traditional human action recognition methods need to extract features by manual design,but the features extracted by manual design have defects which may not contain differential information.In this paper,we use convolutional neural network to automatically extract features to determine the behavior category,and fuse the refinement ideas used in many object classification tasks(which may be different in literal expression).In short,we use the information provided by body parts to determinate the global result.The experimental results show that the recognition accuracy of this method is higher than that of the traditional ways.(2)The Pascal VOC2012 data set and Stanford 40 data set used in experiment lack enough samples and equilibrium.In order to make the experiment procedure more scientific,firstly,the amount of data was expanded to solve the problem.The original sample image is reversed,randomly tailored and deformed to a certain extent to generate new data samples.The generated samples are brand-new samples.As for the imbalance problem of data samples,the sample data is equalized so that each behavior category contains the same amount of samples.