Ground Targets Detection And Tracking From Unmanned Aerial Vehicle Perspective

The detection and tracking of moving targets in aerial video is a key technology in the tasks of Unmanned Aerial Vehicle(UAV)investigation,disaster relief exploration and traffic monitoring,which plays a vital role in the development of the civil sphere and the construction of the military field.The main features of aerial video are relatively small targets,low image resolution,complex background,and relatively small motion of the target relative to the background.These features make researchers face challenges in detecting and tracking targets in aerial video.This thesis mainly studies the method of target detection and tracking from the perspective of UAV.The main content of this thesis are as follows:1.A target detection algorithm based on deep learning of convolutional neural network is studied.The algorithm principles and flow charts of Faster Region Convolutional Neural Network(Faster R-CNN)and YOLOv3 are introduced.Aiming at the deep feature abstraction problem of Faster R-CNN algorithm,different scale feature information fusion is used.Faster R-CNN,improved Faster R-CNN and YOLOv3 are trained and tested on datasets taken by low altitude UAV and small target data sets from high altitude perspective respectively.The improved Faster R-CNN algorithm is superior to Faster R-CNN algorithm in accuracy and processing time.The performance of target detection algorithm based on deep learning is analyzed and the generalization ability of different methods is tested.2.A target tracking algorithm based on correlation filtering is studied from the perspective of UAV.This thesis introduces the algorithm principle of correlation filtering target tracking,gives the principle of training classifier,detecting target and updating classifier,and summarizes the algorithm flow of correlation filtering target tracking.A variety of target tracking algorithms based on correlation filtering are analyzed quantitatively.It is concluded that under the same conditions,it is better to use intensity feature,fHOG feature and color feature fusion to represent the target.Using UAV data sets of aerial photographs,the performance of several target tracking algorithms based on correlation filtering is analyzed and compared,and the advantages and disadvantages of different correlation filtering methods are compared.3.Aiming at the problem of complete occlusion in target tracking,an improved algorithm based on correlation filtering is proposed.Add a re-detection mechanism when encountering complete occlusion during the moving target tracking process,mainly using the motion information of the target,predicting its motion trajectory,and predicting the position of the target in this frame.At the same time,the target template before the occlusion is saved,and then the correlation between the target template and the sample at the predicted position is calculated,and its maximum response is taken to fine-tune the target position.The improved algorithm is compared with the algorithm with re-detection module and the algorithm without improvement on the UAV dataset.The improved algorithm has obvious advantages in accuracy and real-time.