| In recent years,the People's Armed Police starts to take charge of the armed patrol task in key regions in the cities,which brings new serious challenges to soldiers' abilities of dealing with sudden events.Therefore,how to make better use of equipment and technical means to assist the completion of duty tasks become increasingly concerned about the construction of the problem.On the other hand,with the further reduction of UAV costs and the increasingly sophisticated reconnaissance technology,it is highly likely to be widely used in the armed police officers and men on duty,providing new technical support for the completion of the task.This paper studies the target detection and tracking of UAV video,including the dynamic background motion estimation,running pedestrian detection,single target and multi-target tracking,etc.The main work and achievements are as follows:1.Considering the accurate registration of the dynamic background of UAV,put forward the method of predicting the foreground feature points based on the significance of motion and verify the applicability of the algorithm.Firstly,with analyzing the methods of image registration,feature-based registration method is used.Secondly,compare the advantages and disadvantages of the feature extraction operators and select the excellent feature algorithm.And then a foreground feature pre-elimination method based on motion significance is proposed,double screening out the wrong ORB feature points.Finally,the affine transformation model is used to register the adjacent images to obtain the transformation matrix and the global motion is compensated.2.In view of the characteristics of running pedestrian movement,studiesthe feature extraction method based on discrete cosine transform and with the histogram of oriented gradient,identify running pedestrians.Firstly,review the extracting process of histogram of oriented gradient in pedestrian detection.Then,the discrete cosine transform method is introduced to analyze the characteristic curve of running behavior in frequency domain.Secondly,on the basis of introducing support vector machine,use combination features to realize the training and detection of running pedestrians.Finally,through the experiment,it is proved that the combination feature is superior to the single histogram of oriented gradient feature in running pedestrian detection.3.Introduce the adaptive color-studying process to improve the target tracking technique based on color feature,which aims to solve the problem of tracking drift when the target is obscured,or scale and light changes.Firstly,introduce two kinds of trackers based on circulant structure with kernels(gray feature)and traditional color-based feature and analyze the problems.Secondly,it is proposed to use adaptive learning rate and adaptive Gaussian kernel to respectively improve the update and marking methods of traditional color feature tracker training model so as to reduce the target model cumulative error.And the experiment results show that the improved method could solve the above problems.4.In the light of the problem of target-missing when the target is dense or the target color is similar,use the linear programming scheme to improve the multi-target tracking method based on the joint probabilistic data association.Firstly,the basic principle and steps of the traditional joint probabilistic data association method are introduced.Then,the linear programming scheme is used to simplify the joint event association algorithm and to adaptively solve the associated event parameters.Finally,the experiment uses multiple self-timer and public video data sets to verify the effect of improved algorithm including horizontal comparison of different algorithms.The results show that the joint probabilistic data association algorithm based on adaptive solution has good robustness to multi-target tracking. |
PDF Full Text Request