Visual Object Detection And Tracking Technology Of UAV

Visual object detection and tracking technology is the core research direction in the field of computer vision.Unmanned aerial vehicle can provide rich and varied videos for object detection and tracking technology or carry computing platform to realize autonomous detection and tracking of targets.However,the difficulty of object detection and tracking technology is greatly increased due to the jitters of unmanned aerial vehicle,the complex and changeable environment and the occlusion and morphological changes of the target.With the continuous maturity and development of deep learning theory,deep learning plays an increasingly significant role in the field of computer vision.Especially in the object detection,with the successful application of deep network,the accuracy of detection algorithm is constantly improved.This thesis focuses on the combination of the detection algorithm based on deep learning and tracking algorithm to solve many difficulties and challenges in the practical application process.Although researchers at home and abroad have achieved fruitful achievements on the application of deep learning in the field of object tracking,there are still some problems need to be perfected: 1)In terms of the real-time performance of the algorithm,due to the huge parameter calculation of the deep network,the object tracking algorithm based on deep learning usually fails to meet the real-time requirements;2)In terms of the target correction,when the tracking algorithm fails,most tracking methods cannot recover robust tracking of the target;3)In terms of the target conversion,when the target category in the tracking process needs to be changed in time,most of the existing methods can not achieve the autonomous conversion of the target;4)In terms of the multi-object tracking,traditional tracking algorithms are usually unable to deal with the emergence of new targets and the disappearance of old targets.The existing algorithms propose a series of solutions,but the computational complexity is too large.In view of the above problems,this thesis mainly carries out research from the following three aspects:(1)Going into the traditional and deep learning-based object detection and tracking algorithms.In this thesis,the traditional detection and tracking algorithms and deep learningbased detection and tracking algorithms are described in detail in terms of the principle,process descriptions,formula derivation and experimental comparison.The characteristics of different detection algorithms and application scenarios are summarized to provide the concrete reference basis for the algorithm selection in practical applications.(2)The fusion of deep learning-based detection algorithm and single object tracking algorithm.In this thesis,we firstly retrain the network weights of YOLOv4 by purposely selecting the dataset so as to meet the needs of practical application better.Then,we combine the high accuracy of YOLOv4 based on the deep network and the real-time performance of the traditional tracking algorithm.And we innovatively achieve the goal of automatic initialization,tracking correction,automatic conversion.Finally,the proposed algorithm is verified to have better real-time performance,autonomy,robustness and scalability in specific scenarios by using representative experimental videos such as unmanned aerial vehicle photography.(3)Improving the multi-object tracking algorithm’s performance by YOLOv4.Firstly,we introduce the systematic process of the traditional multi-object tracking algorithm and the multi-object tracking algorithm based on deep learning.Then,we analyze the difficult problems in multi-object tracking technology,including the problem that the traditional algorithm can’t solve the change of the number of targets in the tracking process and the problem that the speed of the algorithm based on the deep network is too slow.Finally,a multi-object tracking algorithm with high accuracy and robustness is implemented by YOLOv4 algorithm for object detection,Kalman filtering algorithm for target position prediction,and Hungarian algorithm for target inter-frame matching,and the performance of the algorithm is proved by experiments based on aerial footages of the unmanned aerial vehicle platform.