Research On The Visual Detection And The Bahavior Recognition Technology Based On Deep Learning

The visual detection and the behavior recognition are important research contents in the field of computer vision.In recent years,the visual detection and the behavior recognition based on the deep learning have been widely used in the transportation and the industry and the commerce.This paper analyzes several commonly target detection and behavior recognition algorithms and studies the Mask RCNN target detection algorithm and the C3 D behavior recognition algorithm based on the deep learning.At the same time this paper optimizes their shortcomings and proposes the improved algorithms.The specific work is as follows:(1)This paper introduces the target detection and the behavior recognition technology and presents the principle of the target detection based on the deep learning.The basic theory of the convolutional neural network and the full convolutional network and the gradient descent method of the deep learning and the target detection data set are introduced.The principle of the behavior recognition and detection based on the deep learning is discussed.The traditional behavior recognition algorithms and the behavior recognition algorithms based on convolutional neural networks are analyzed,as well as Faster RCNN and Mask RCNN algorithms based on the regional recommendation target network.The dual-stream convolutional neural networks and the 3D convolutional neural networks and the deep learning frameworks are introduced in the network behavior recognition algorithm based on the convolutional neural.(2)This paper proposes a vehicle damage area detection algorithm based on improved Mask RCNN.The convolutional neural network is used to detect vehicle damage areas.The training data sets is made according to the feature attributes of the detection target.The target candidate regions are generated by the region suggestion network.The residual feature pyramid network is used to extract the feature map.The anchor frame ratio and the threshold of the region suggestion network are adjusted.The non-maximum values Suppression algorithm is replaced by the Soft-NMS.The ROI classification and the bounding box regression and the mask generation are used by the ROI Align technology.The convolutional network is trained and the final optimized detection model is obtained.(3)As a component of the research about the behavior recognition,the C3 D algorithm which is learning the spatio-temporal features is easy to lose important feature.For this reason,this paper designs a dual-stream-based 3D convolutional neural network.On the spatial stream,the 3D convolutional neural network is used to extract the content about video and the information of the characterization.In the time stream,the TVNet optical flow algorithm is introduced.The video optical flow data are extracted through the TVNet algorithm and the motion characteristics and the time characteristics of the video are extracted.The feature fusion method is used to fuse feature vectors and the prediction results are obtained by the linear SVM.The experimental results show that the network architecture proposed in this paper can improve the behavior recognition rate in the space and time.