Research On Gesture Recognition Method Based On Computer Vision

With the development of modern technology and the rapid improvement of computer performance,human-computer interaction has gradually become naturalized and diversified.Therefore,it is of great significance to study flexible and natural human-machine control methods,especially relying on gesture recognition.If you use computer vision to perform gesture recognition instead of using hardware devices such as data gloves and remote controls,it will greatly facilitate people's needs.Computer vision can realize simulation of human vision by using algorithms.It can be applied to control,medical image analysis,target detection,etc.It can also be applied to many fields such as artificial intelligence,image processing and scientific computing.the ultimate goal is to make the computer can be like Humans see and understand realistic scenes through vision.Allowing computer vision to understand the information conveyed by people's gestures can reduce the constraints imposed by contact devices.Gesture recognition is a more challenging technique.Traditional image processing is difficult to achieve excellent results because the deformation of the human hand is uncertain,and the target positioning,motion state,motion sequence,environmental interference,and easily obscured by objects are also considered.The confusion of various problems,which makes the recognition of gestures become the focus and difficulty of long-term research in the field of computer vision.Based on the deep learning Tensorflow framework and Ubuntu16.04 operating system,the gesture recognition is realized in two ways.The simulation results show that the proposed improved algorithm achieves good gesture recognition.The main research work completed in this paper is as follows:(1)The basic theory of convolutional neural networks is expounded,including the principle of convolutional neural networks,the characteristic output of convolutional and deconvolution layers,the optimization of objective functions and the principle of residual techniques.It lays the theoretical foundation for the research and implementation of subsequent gesture image recognition algorithms.(2)The gesture image segmentation and recognition method based on improved capsule network and algorithm is proposed.The gesture recognition is realized by gesture segmentation,gesture positioning and gesture recognition.Firstly,the improved capsule network algorithm is combined with the U-net network and residual technology to realize the U-shaped residual capsule segmentation network,and the segmentation of the gesture image is completed.Then,the structure of the matrix capsule network is optimized,and the recognition of different types of gesture images is completed;Finally,through the training and testing of dataset,the performance index,P-R curve,ROC curve,the visualization of middle network layer feature,gesture image recognition confusion matrix,gesture image recognition loss curve and recognition accuracy curve of gesture image segmentation are obtained.The simulation results show that the proposed improved model achieves better recognition results.(3)The method of multi-model and algorithm synthesis based on Tiny Yolo v3,Deep-sort and DenseNet is studied to realize gesture recognition.The specific research analyzes the Tiny Yolo v3 detection algorithm,the Deep-sort target tracking algorithm and the DenseNet algorithm.By making the data set and training the model,firstly train the three models Tiny Yolo v3,Deep-sort and DenseNet,then these three models and algorithm is integrated to realize a method for detecting,tracking and recognizing gestures in a complex background.The algorithm is up to7~8 frames per second.