Research On Car Washing Behavior Recognition Method Based On Deep Learning

Behavior recognition is the focus of computer vision and intelligent monitoring.At present,the recognition of human behavior in simple context has achieved high recognition rate,but there are still shortcomings in the research of behavior recognition under complex field.In this thesis,the car wash line of this complex environment as the background of human behavior recognition research.The traditional behavioral recognition algorithm is characterized by artificial design,which is not only time-consuming but also choose a good feature depends on the experience and luck while adjust it takes a lot of time.Deep learning abandons the traditional method of relying on artificial design features,through building a multi-layer neural network to achieve the machine automatically learn to hidden the relationship within the data.This approach makes the learning behavior characteristics more accurate and conducive to recognize the behavior correctly.As a typical network in depth learning algorithm,convolution neural network has achieved good results in the field of image,but there are still some shortcomings in the recognition of video-input behavior,which need to be improved.This thesis introduces the 3D CNN that can be accepted as the input of the network and is applied to the behavior recognition.It extends the two-dimensional convolution operation to the three-dimensional so that the network can not only learn the static image content learn the movement information of the human body in the video.But the down-sampling layer of the network is still two-dimensional down-sampling,and can not accept frames with different resolution different video input.Based on the shortage of 3D CNN network,this paper makes four improvements to its network structure.The nonlinear MLP convolution operation is added to the convolution layer,which makes the network more abstract.Because the time domain and space domain also have certain invariance,so the down sampling layer will be extended to three-dimensional,joined the time domain down sampling,while preserving useful information to reduce the network need to learn the parameters to improve network performance.And then joined the space-time pyramid down sampling technology,not only avoid the loss of input information,but also allows the network to accept different frame length and resolution of the video input.The ReLU nonlinear function is used to replace the tanh function of the original network as the activation function,ReLU function is not saturated function,not like tanh tangent function in the training to reduce the back propagation error,but also can accelerate the training network convergence,Finally,the softmax classifier is adopted.The input of the improved network will discard the gradient channel of the original network and let the network automatically learn the gradient information from the data samples.Finally,experiments were carried out on KTH datasets with simple background and complex background data sets,and the experimental results were given.The advantages of the improved algorithm and the influence of the data sets on the experimental results were analyzed.The space-time complexity of the network is an important index to evaluate the network performance.At last,the space-time complexity of the improved network is analyzed in detail.