Static Gesture Recognition Based On Deep Learning

With the development of smart devices,gestures,as a means of human-computer interaction,are rich,flexible,and intuitive.The use of gestures as a means of human-computer interaction is more in line with people's living habits.Gesture recognition is of great significance in helping deaf-mute groups integrate into normal social life,machine control,and safe driving.Among them,gesture recognition technology based on external devices such as data gloves and Kinect has been relatively mature and widely used,but the gesture recognition technology based on computer vision is still not mature enough.Traditional image processing methods usually choose to study the design of algorithms in a simple or single background in order to avoid environmental influences,which lead to the low practical value of the algorithm.In the complex background,this paper aims to improve the speed of gesture detection and detection accuracy,and adopts different methods to study and improve gesture recognition.Through experimental comparison,the problems and effects of various algorithms in gesture recognition are analyzed.The final improved YOLOv3 deep learning model can achieve real-time detection of gestures under the premise of ensuring the accuracy of gesture detection,and has good application value.This paper mainly uses three methods to study gesture recognition:(1)For traditional image processing algorithms,it is impossible to effectively segment the skin color,such as arm and face.The traditional sliding window detection method generates a large number of window images and affects the processing speed of the algorithm.The image pyramid structure is constructed according to the custom image scaling rules for the skin-like area,and the window image is generated in combination with the sliding window operation to perform target recognition and classification recognition for each window image.This improvement makes the algorithm processing speed reach 25fps./s,the usual range is expanded,gesture recognition accuracy reaches 61.2%,and the detection accuracy is not high.(2)Using the Faster RCNN deep network model for gesture recognition,the use of Faster RCNN and FPN algorithm is combined with Faster RCNN to improve the detection accuracy of small targets.The detection accuracy is as high as 97.6%,but the detection speed is very slow(2fps/s),and real-time detection cannot be achieved.(3)The YOLOv3 deep network model is used for gesture recognition research.For YOLOv3,there is an inaccurate positioning of the target frame for close-range gestures,and the performance is degraded when the IOU increases.It is proposed to re-fine the experimental data by k-means,generate the size of the a priori frame adapted to the dataset of this paper,and remove the prediction of the target in the top-level feature map to improve the positioning accuracy of the target gesture.In the end,the detection accuracy of the model reached 98.6%,and the detection speed was 22fps/s,which can achieve the effect of real-time detection.