Real-time Compressive Tracking Based On Transform Domain

With the rapid development of computer vision and artificial intelligence recently,there have been great advances in object tracking task.Currently,object tracking has been widely applied in human-computer interaction,intelligent transportation,security monitor-ing and navigation detection and so on,its broad market prospects and important application values attract more and more attentions from industry and academia.To improve the perfor-mance of object tracking,researches attempt to explore some machine learning approaches to construct a real-time,robust,accurate tracker.However,although the tracking perfor-mance has been greatly improved,there are still many practical problems to be solved,such as the effect of occlusion,illumination variation,pose change,rotation and scale change.Among the research topics in object tracking task,it is crucial to find a reliable and effective appearance representation for object.This paper focuses on the appearance rep-resentation as well as the performance associated with it.We propose object tracking al-gorithms based on domain-transform,including two works.The first work aims to solve the tracking problem under some complex circumstances such as occlusion,illumination and pose variation.We propose a tracking algorithm integrating Fourier Transform and template matching strategy.The algorithm integrates both history and latest appearances with an adaptive weight to generate an object template.Moreover,the object template is transformed into sparse frequency domain by Fourier Transform so that the original image information is greatly compressed.As a result,the approach construct a compressed and robust object template to achieve a high tracking accuracy with low computational cost.Based on the first work,the paper further propose a Nonsampled-Countourlet-Transform based on tracking algorithm for solving the insufficiency of Fourier Transform in orien-tation and scale analysis for object.The approach extracts multiple scale and orientation information for the object in original image,and then generates a compressed feature vec-tor by using the compressive perception technique.The feature vector is finally fed into Adaboost classification algorithm to construct a robust classifier for the tracking task.To prove the effectiveness of the proposed two approaches,we make an experimental comparison and analysis on two classical object tracking video datasets.It is demonstrated that the Fourier-Transform based object tracking algorithm could achieve a real-time track-ing with high accuracy under the circumstance of illumination variation,occlusion and fast movement.However,the performance is limited when the object is rotated or stretched.On the other hand,the Nonsampled-Countourlet-Transform based tracking algorithm relieves the rotation and scale change problem,and show a significant improvement as compared to the state-of-the-art approaches.