Research And Improvement Of Visual Trcking Algorithm Based On Kernel Correlation Filter

Recent year witness the rapid development of computer version.Visual object tracking, object recognition, object detection and classification become research hotpots in this field. As a sub-topic of computer version, object tracking is far away from apply to the practical work, for the reason of its opening and complication. There are many factors that make a great challenge to the tracker. Such as the complicate trajectory of moving object, object appearance change, partial or fully occlusion, illumination variety. Over the years, the correlation filter based trackers have aroused much attention due to its robust and effectiveness.In this paper our goal is to review the correlation filter based trackers,analysis its theory and take additional methods to modify its shortcoming.The rest of the paper is arrange as follows:1?In the beginning, we will explain some theories about correlation filter based tracker, including feature extraction, training examples selection, fast learning and detection, kernel tricks. This paper will expound how this stuff work together forming a correlation filter based tracker.2?What follows is how we tackle two main problems exit in correlation filter based trackers, namely scale variety and occlusion. We introduce a multi-scale model to handle object scale change, and an adaptive model update strategy to resist the effect of object occlusion.Experimental result verify that our method achieve promising result.3?This paper studies the framework of tracking algorithms in the general sense, including feature extraction, motion model, observation model, template updating and reinforcement learning. In this paper, we use the popular depth convolution neural network to characterize the target, and combine it with the existing correlation filter framework, the tracking effect is greatly improved.