Human Action Recognition Based On Convolutional Neural Network

With the continuous improvement of people's life,the research of vision-based human behavior has been widely used in the fields of intelligent transportation,smart medical,video surveillance and human-computer interaction,which has gradually become a research hot spot.However,due to the complexity of human behavior,the diversity of background environment,and the change of camera angle,the traditional methods of recognition human behavior are prone to problems such as huge workload,and low recognition accuracy.In recent years,deep learning has provided a new research direction for the research to recognition human behavior along with the development of artificial intelligence.As a branch of deep learning,convolutional neural networks do not need to artificially extract features,which use neural networks to simulate human brains and automatically learn and extract image features to make up for the shortcomings of traditional methods.Therefore,this paper designs a method of recognition human behavior based on deep learning,which used the advantages of convolutional neural network and combined with foreground detection algorithm and convolutional neural network algorithm.Firstly,due to the LeNet-5 network gaining low recognition accuracy in complex background,this paper proposes an improved convolutional neural network model.By increasing the number of network layers and convolution kernels,the network can extract more target features.It has achieved good recognized results in image classification and human behavior recognition.Secondly,in view of the problem that the traditional foreground detection algorithm has poor performance in the complex environment,this paper improves it.To better handle precampus attractions,it use a threshold associated with the model sample;And it proposes to use the skeleton information instead of the traditional foreground image by using the Kinect camera to collect body skeleton information.It solves the environmental factors such as occlusion and illumination.Finally,according to the time series of human behavior,a continuous frame combination method of behavior recognition by using the average of the predicted probabilities of successive image sequences as the final result.It is proposed to avoid the interference between different behaviors which have similar actions and improve the accuracy of the model.In term of the influence of network model parameters on the final recognition rate,this paper compares different activation functions,optimization algorithms and dropout coefficients,and selects the model parameters that are most suitable for the experiment in this paper,which improves the recognition accuracy of the model.The experimental results show that the human behavior recognition method based on deep learning has achieved good results in motion foreground detection and behavior recognition.