Infrared Target Tracking Based On Kernelized Correlation Filter

Infrared target tracking is widely used in military,transportation and security.There is still a big challenge to the tracking algorithm due to the factors such as low resolution,target occlusion,similar target,camera movement in complex scene.The thesis studied the tracking of complex ground infrared target,the main achievement of this dissertation is as follows:First of all,we introduce the kernelized correlation filtering algorithm.Kernelized correlation filter is a fast tracker and consumes less computing resources.But there are some defects in the expression of target features.Kernelized correlation filter is easy to lose the target when tracking low resolution targets,because of fewer features.We combine the gray feature with the HOG feature to solve the problem and achieve a good result.Secondly,in order to tracking the target in occlusion and similar objects overlapping successfully,we combine the particle filter with the improved algorithm described above,a kernelized correlation filter based on position prediction and combined features is proposed.When target is in the occluded state,the apparent information of the target is obscured,trackers based on the appearance features is easy to lose the target.Therefore,we use the particle filter to estimate the target position and motion information,and then use the improved kernelized correlation filter to accurately calculate the target position.The combined tracker has better anti-occlusion ability.Thirdly,the scale pyramid and gray distribution histogram is used to estimate the target scale.This is a real-time approach,and it can estimate the target scale adaptively.We use the adaptive learning rate to update the target model,so the model adapts to the target appearance change.At last,we collect and mark a number of infrared video sequences and build an infrared video dataset.The algorithm proposed in this paper is compared with the classical tracking method in the infrared video dataset.The algorithm is evaluated and analyzed from the aspects of precision rate,success rate and time complexity.An algorithm debugging platform is coded and the transplantation of algorithm is facilitated.