Research On UAV Ultra-low Altitude Head-up Target Tracking Technology Based On TD3

With the increasingly advanced drone-mounted sensor technology and flight control technology,UAV are widely used in military and civilian fields as a convenient and efficient equipment for data acquisition and detection.However,in some scenarios,common UAV overhead tracking or Pan-Tilt tracking schemes are difficult to collect target information.This thesis analyzes the following two technical difficulties for a new application scenario of a common UAV equipped with a monocular camera for ultra-low-altitude head-up tracking targets:the dramatic and quick transformation on the scale and position of the target in the video collected by the UAV camera during ultra-low-altitude tracking.This thesis proposes a video stream recognition algorithm based on the KCF filter algorithm.The algorithm uses the YOLO to solve the problem that the KCF cannot perceive the scale change of the target in the video stream.At the same time,this thesis designs the HOG feature recognition direction neural network,the frame buffer system of YOLO and KCF,and the mechanism of the experience buffer to adapt to the characteristics of highspeed and large-scale changes in the target scale and position in the video when tracking the target at ultra-low altitude.After obtaining the position and scale of the target in the frame through the KCF optimization algorithm,this thesis proposes an intelligent low altitude tracking algorithm of UAV Based on TD3（Twin Delayed Deep Deterministic policy gradient algorithm）as the tracking strategy of UAV.Based on the MDP process modeling of the UAV tracking process,the concepts of state,action and reward function in the reinforcement learning algorithm are designed at length.The algorithm proves the excellent effect of the UAV head-up tracking at low altitude in the follow-up experiments.In this thesis,the Tello UAV is selected as the experimental drone,and the UAV ultralow altitude head up tracking target scene is built.Based on this scene,the UAV tracking system is designed and developed,and the video stream recognition method based on KCF is applied to this system.Compared with the YOLO recognition algorithm using the recognition speed as an indicator,it proves that the recognition algorithm proposed in this thesis can be adapted to the situation where the target scale and position change rapidly and greatly during the ultra-low altitude tracking process.In order to verify the intelligent tracking algorithm of UAV Based on TD3,a simulation environment is developed by using gym platform as the environment platform and box2 d as the physical engine to simulate the process of UAV tracking targets.At the same time,a series of benchmarks are proposed to evaluate the drone motion tracking strategy.This thesis compares the proposed tracking algorithm with other tracking algorithms based on deterministic policy reinforcement learning.The results show that the tracking algorithm proposed in this thesis has the best tracking effect.