Research On Tracking With Multi-Kernelized Correlation Filter For Object Following Of Unmanned Aerial Vehicle

At present, intelligent flight technology is the hot topic in the field of Unmanned Aerial Vehicle(UAV), which has attracted the scholars to concentrate and research it deeply. During the following task, UAV utilizes the forward-facing camera to locate the target, estimate motion state, and then turn to control its flight trajectory from the image feedback to follow the target. With the widely application of UAV in shooting of extreme sports such as athletics, skiing, rowing, etc. and recording daily life, object following for UAV has important significance and development prospects. But the exploration and research related to following is still in the exploratory stage. The key technology of object following utilizes the vision-based detecting and tracking of arbitrary objects, which is the basis to follow safely and accurately. This paper mainly focuses on the visual target detecting and tracking. The main work is as follows:Firstly, the existing algorithm of moving object detecting and tracking is reviewed, such as optical flow, background-difference, inter-frame difference and object tracking based on correlation filter, of which the mechanism and performance are analyzed. Then the tracking method of fast speed and high accuracy, high-speed tracking with kernelized correlation filters(KCF), is studied. In this method, the dense sampling is accomplished by cyclic matrix, and the computation of the kernel function matrix and the training of the classifier are accelerated with its property. After that, the experiment is conducted at the aspects of the complexity of the algorithm, the change of illumination, occlusion and attitude variation. The results show that this method has a good performance. But when the object is occluded, tracking fails.Secondly, aiming at the disadvantage of KCF, a cascaded iterative classifier initialization method is proposed to settle the problem of classifier initialization with single sample. An adaptive classifier is designed to obtain a strong classifier, which utilizes the integration of weak classifiers. The experiments demonstrate that the integrated strong classifier contains the target information in the multi frame, and has the ability to detect the mutative target.Then, this paper proposes a multi kernel parallel correlation tracking algorithm in order to solve the problem that is difficult to deal with the problem of occlusion in the trackingalgorithm based on kernel function correlation filtering. It can discriminate real target and false target accurately, thus ensuring the accuracy of classifier. In this method, the position of the target is estimated by three classifiers, and the classifier is updated according to the output response. Besides, this multi classifier use parallel detection and is update independently. So it can not only ensure the classifier adapt to the variety of targets, but also prevent the loss of real target information caused by the integration of false target information.In order to verify the effectiveness of the proposed algorithm, this paper conducts a set of comparative experiment through eleven videos. The results show that under the same conditions, the processing speed of improved algorithm(79fps) is slightly slower than KCF(115fps), but its accuracy rate increases from 78 percent to 87 percent, while the speed of Struck which is the state-of-art algorithm is only 8fps?In target occlusion test, the improved algorithm is more robust than the original algorithm and can track the target accurately when fully occluded, while the latter loses the target.