The Gesture Recognition Research Based On The Improved Faster RCNN Algorithm

In modern industry,robots can bear huge workload,have high repetition ability and productivity.At the same time,robots can provide the strength required for handling heavy objects and precision required for other operations.Therefore,robot technology is widely used in industry.The human-computer separation lacks flexibility and safety,which can no longer meet the needs of production and life.Therefore,the research of human-computer interaction technology is of great significance for industrial production and life.Gesture is used as an effective communication method in human-computer interaction.There are two kinds of gesture recognition methods: gesture recognition method based on external devices and gesture recognition method based on computer vision.Gesture recognition method based on external devices has been relatively mature and widely used in virtual reality,sign language recognition and robot production.However,gesture recognition based on computer vision has to be improved.Therefore,based on the interaction between human and industrial machines,this paper proposes a gesture recognition method based on improved Faster RCNN.And the main work is as follows:(1)In order to solve the problem that the traditional computer vision algorithm needs artificial feature extraction for gesture recognition,convolutional neural network algorithm is used to extract features automatically to detect and recognize ten kinds of gestures.Compare the pre-processing effects of mean filters and Gaussian filters with different convolution kernels on the gesture images,and select Gaussian filter to pre-process;build convolution neural network model,set parameters,train convolution neural network,and use the model to detect and recognize ten kinds of gestures.The accuracy of recognition results is 97.5%,which indicates that gesture recognition method based on convolutional neural network algorithm can extract features automatically,and avoid the problem of low recognition accuracy caused by manual feature extraction.(2)Aiming at the problem of low accuracy and robustness of convolutional neuralnetwork for gesture recognition,the Faster RCNN algorithm is used to detect and recognize ten kinds of gestures.Gaussian filter is used to pre-process the gesture images in the NUS gesture data set;VGG16 and residual network are used to extract the features of the gesture images respectively;Five fold cross validation is used to improve the generalization ability of the model.The accuracy of recognition results is 99.89%,and the accuracy of gesture detection and recognition has been improved significantly,which is suitable for practical industrial human-computer interaction applications.However,the accuracy of the small gesture recognition of the algorithm needs to be improved.(3)In view of the low accuracy of the Faster RCNN algorithm in detecting and recognizing small gestures in images,the Faster RCNN algorithm combined with feature fusion is used to detect and recognize small gestures in the images.Gaussian filter is used to pre-process the gesture images;different sampling strategies are used according to the different levels of convolution neural network: for the shallow feature map,the pooling operation is used;for the deep feature map,the deconvolution operation is used.At the same time,according to the importance of each feature map to the recognition task,the weight of each feature map is automatically assigned.The accuracy of gesture recognition is increased to 94.25%.In this study,convolutional neural network and Faster RCNN algorithm are used to achieve accurate gesture recognition,which provides a feasible method for human-computer interaction in industry,and it is of great significance for robot production,smart home,sign language recognition.