Research On Single Target Recognition And Tracking Algorithm For UAV Based On Deep Learning

With the popularity of artificial intelligence,traditional industries are constantly changing.As far as UAVs are concerned,more and more companies or organizations are expanding their applications with tools such as computer vision or deep learning.Due to the limitation of the computational performance of the flight control processor,the traditional drone usually needs to encode the captured image and transmit it to the ground for image processing.However,when the image is transmitted to the ground,the work of encoding and decoding will consume extra time,which affects the realtime performance of the entire system.In addition,in the complex electromagnetic environment information transmission of the drone and the ground station will be greatly disturbed,which seriously affects the stability of the system.This project adopts a robotic operating system and deep learning algorithm on the UAV side to perform real-time processing of camera video through computer vision,in order to realize real-time recognition and tracking of the target by the UAV.Among them,the UAV target recognition algorithm is implemented based on MobileNets neural network,and the UAV target detection algorithm is implemented based on MobileNet-SSD neural network.The target detection algorithm is used to obtain the coordinates of the target center point.Based on the principle of pinhole imaging,the accurate coordinates of the target in the world coordinate system are calculated by linear transformation between the relevant coordinate systems.However,the linear model does not consider the lens distortion during the calculation process,so the accuracy will be slightly wrong.And the camera calibration method will be used to calculate more accurate coordinates.After obtaining the accurate coordinates,the UAV flies from the current coordinate to the next coordinate point through the coordinate positioning of the GPS to realize the tracking of the target by the drone.Based on ROS and Gazebo,the UAV simulation verification platform was built,and the UAV was debugged and calibrated by QGroundControl.The above single target recognition and detection algorithms were verified and analyzed by hardware-in-the-loop simulation.The single target recognition algorithm of MobileNets neural network and the single target detection algorithm based on MobileNet-SSD neural network can achieve better results and meet the relevant requirements in this subject.