Human Behavior Recognition Based On Convolutional Neural Network

With the popularity of video devices,human behavior recognition technology is widely used in the many fields such as intelligent manufacturing,intelligent monitoring and so on.Traditional human behavior recognition methods usually perform behavior recognition through artificially designed features,such as template matching and 3D modeling.Although these methods can achieve a better recognition rate in a simple video of a scene,the recognition effect is unsatisfactory when faced with a scene with a complicated scene in reality.In the human behavior recognition task,the main difficulty is to accurately extract the human body's motion characteristics and time domain features,and the artificially designed features are difficult to do.With the development of convolutional neural networks,the way of artificially designing features is gradually replaced by the method of automatically extracting features from convolutional neural networks.A large number of image recognition tasks and video detection tasks are solved.Therefore,this paper proposes a human behavior recognition method based on convolutional neural network.(1)Human target detection.In the video,a human body target detection model was designed for the human body.Based on the target detection model of YOLO-V1,the model first replaces the main network architecture in the original model of YOLO-V1 with resnet18,which reduces the weight of the model and improves the detection speed of the model.Next,the output of the model is modified to change the multi-target detection in YOLO-V1 so that it only detects human targets.Finally,the loss function of the model is re-optimized,and the regularization loss term is introduced to improve the robustness of the model.(2)Human body posture detection.Based on the detection of human target,a new residual unit whose name is the cross residual,is designed,which is used to improve the residual structure in the Stacked Hourglass human pose recognition model,so that the model has less weight.The amount of calculation is smaller.In addition,the improved human body posture recognition model incorporates the shallow features of human target detection,which further ensures the recognition accuracy of the model.(3)Human behavior recognition.On the base of human target detection and human body gesture recognition,the cyclic neural network is used to extract the time domain features in human behavior recognition.First,based on the cyclic neural network of the LSTM structure,the input gate in the LSTM structure is improved,and the sigmoid layer in the input gate is removed.Next,the human body joint point information of each frame in the video is extracted by using the human body gesture recognition model.Finally,the position information of the human bone joint points is used as the input of the improved LSTM structure cyclic neural network to complete the human behavior recognition.In addition,the highdimensional feature vector in the human target detection is merged with the human bone joint point information to further improve the accuracy of human behavior recognition.Experiments show that the human behavior recognition method based on convolutional neural network designed in this paper achieves good accuracy on NTU RGB+D dataset.