Research And Application Of Gesture Recognsition Algorithm Based On UAV

With the rapid development of UAV technology,UAV have been widely used in human production and life.How to interact with UAV conveniently,quickly and simply has become a hot spot of current research.There are many ways of human-computer interaction with UAV,such as voice interaction,human posture interaction,face interaction,gesture interaction and so on.Gesture interaction is natural,convenient and intuitive,which is more suitable for interaction with UAV.From the perspective of application fields,the interaction between gestures and UAV has great application value in the take-off and landing,position adjustment and visual tracking.In this paper,a human-computer interactive system is designed.Firstly,gesture recognition algorithm is run by processor.Then the video stream obtained by the monocular camera is recognized and the gesture classification and location information is output.Finally,the classification results are converted into command to interact with UAV.Due to the complexity of the environment,the diversity of gestures and the lack of computing power of embedded devices,the accuracy and speed of gesture recognition are seriously affected.For the above-mentioned problems,this paper improves the accuracy and speed of gesture recognition by improving gesture recognition algorithm,and designs an interactive system of gesture control UAV based on the algorithm.The main content and contributions of this article are as follows:(1)A scheme for interacting with the UAV is proposed.The gesture is recognized through a gesture recognition algorithm,and then instructions are generated and transmitted to the UAV to control its flight status.In addition,for the training and testing of the gesture recognition algorithm model,a dataset of 13,464 gesture pictures was constructed.(2)Aiming at the problem of low recognition accuracy of the YOLOv3 algorithm,an algorithm improvement was made.By using linear stretching anchors(LSA)and adjustablearctangent linear units(ALU),the accuracy of gesture recognition can be improved.After simulation,compared with the YOLOv3 algorithm,the application of LSA and ALU activation functions can increase the m AP value of gesture recognition by 5.8% and 14.8%,respectively.(3)Due to the limited computing power of the processor on UAV,LNNA is proposed to improve the speed of gesture recognition.After simulation,the processing speed of LNNA on2080 TI GPU is 48.9% higher than that of YOLOv3,and the image processing speed on NVIDIA Jetson TX2 has increased from 1.83 frames per second to 11.67 frames per second.At the same time,in order to fully verify the effectiveness of LSA and ALU,simulations are performed on the LNNA network architecture.the application of LSA and ALU activation functions can increase the m AP value by 9.9% and 20.5%,respectively.(4)Transplant the LNNA network to two different styles of DJI drones for humancomputer interaction experiments.Firstly,a visual interface system was developed based on the Tello UAV,which can easily display gesture recognition information and commands information,and send commands to UAV.Secondly,we conducted outdoor flight experiments based on the M210v2 UAV.The whole flight experiment was carried out by ROS system.Through experimental analysis,in close range,gestures can interact well with UAV.