Research On Autonomous Real-Time Target Tracking Method And Application For Rotary-Wing UAV

With the development and progress of science and technology and society,a variety of robots,including drones,are increasingly widely used in all aspects of our daily production and life.The aerial robot system that uses a rotary-wing UAV equipped with a camera as its vision module for target tracking task,can liberate labor and even achieve higher efficiency in many task scenarios,which has been widely used in: disaster rescue,warehouse management,security inspection,event broadcast,military strike and other fields.There are some difficulties in object tracking task,such as the incompatibility of the complex algorithms with the platform,and unstable tracking effect in actual tasks.In view of these difficulties,this paper proposes a object tracking algorithm based on deep learning to achieve autonomous real-time tracking of moving targets on the rotor UAV platform.The specific research work of this paper is as follows:(1)The relevant software and hardware systems are built firstly,and then the transformation model between different reference coordinates is constructed for the hardware main body quadrotor drone,and the kinetic model of the quadrotor drone is also modeled.(2)The overall framework and process of the system are designed into three modules,which are object detection module,object tracking module and tracking control module respectively.The object detection module detects and obtains the position information of the object to be tracked in the initial frame,and then transmits it to the object tracking module for inter-frame tracking of the specified object.As the object tracking module obtains the given initial frame and the position information of the object to be tracked,it locates and tracks the object in the subsequent image sequence,and then feeds back the output of the object tracking algorithm to the tracking control module to complete the autonomous tracking task.According to the control strategy written in advance,the tracking control module issues different control commands to directly control the flight status of the quadrotor drone,and finally complete the object tracking task.(3)In order to balance the accuracy and speed of object detection,an improved object detection algorithm based on YOLO v4-tiny is proposed to provide the position information of the object in the initial frame.This algorithm,on the basis of YOLO v4-tiny,performed a large-scale lightweight operation on its network structure.In addition,attention structures and two topdown FPN modules are utilized to improve the detection accuracy of small objects in the object tracking task of rotary-wing UAVs.(4)Aiming at the problem that the current rotary-wing drone cannot track spacil object in real time,accurately and robustly based on the image sequence,this paper proposes an improved Siamese neural network based on the lightweight neural network Mobile Net v2,which can run in real time on the low-computation on-board computer of the rotary-wing drone.The lightweight structure with less parameter in Mobile Net v2 is integrated into the backbone feature extraction network of the proposed algorithm.What’s more the backbone feature extraction network of the proposed algorithm reduces the network depth slightly compared to Mobile Net v2.At the same time,group convolution,channel reorganization and attention network structure are applicated to improve the feature extraction capability of the network.In order to improve the robustness of the network when the background contains semantic objects,a cascaded RPN network structure is proposed,and feature descriptors with different depths which owe the same spatial resolution are used as the input those single RPN modules,which can fully utilize the extracted deep level features and shallow level features.