Gesture Recognition Research Based On Monocular Vision And Compressed Convolution Neural Networks

The relationship between human and machine has turned from "human adapting machine" to "machine adapting human" with the development of artificial intelligence?computer vision and telecommunication technology.Human-machine interaction system based on gesture recognition has been hot research orientation owing to the improvement of processor performance and the development of advanced gesture catching facilities.However,the diversity of gestures,complex environment background and uncontrollable lighting conditions in the actual scene all pose great challenges to the robustness of gesture recognition.To improve the accuracy of gesture recognition,this paper realizes a multi-strategy identification method that combines visual information and compressed convolutional neural network and designs the gesture recognition control system which aims aerial robots successfully,also makes nature admin interface possible inside which the direct arm gesture interacts with the system.This paper mainly accomplished the following jobs:in gesture design section,taking into consideration of several factors such as the distance between human and machinemovement of the camera and illumination,this paper designs arm gestures adapting to aerial robots;To reduce the economic cost and overload,image acquisition section is designed based on monocular vision.The input of the gesture recognition control system is color image;Target tracking section analysis fast execution,accurately located discriminative scale space tracker(DSST)and kernelized correlation filters tracking(KCF).This section also realized an adaptive switching mechanism for automatically initializing a target tracker using face detection,skin color detection and manually initializing a target tracker;The gesture recognition section,in order to put neural network on aerial robots,merging with deep compression algorithm,improves the speed of network operation while efficiently cut the storage space of weights;Commands generation section designs a state machine for storing multi-frame recognition results in order to improve the reliability of control orders and enhancing the robustness of the system;Algorithm merging section,in order to highly efficiently handle the communication among above sections,based on robot operation system(ROS),realizing the gesture recognition system.Finally,the performance analysis of the system based on Intel(R)Core(TM)i5-6200U and Intel(R)Atom(TM)Z8300/Z8350 two processors so that ensures that the gesture recognition system designed in this paper can be transplanted to an aerial robot.