Research On Multiple Appearance Models Based Long-term Visual Object Tracking Algorithm

Visual object tracking is one of the fundamental problems and also one of the most challenging tasks in computer vision.In recent years,visual object tracking is developing rapidly with the development of some domains such as driverless motoring,visual surveillance,augmented reality and human-computer interaction.Visual object tracking still faces great challenges because the target could undergo significant deformation,heavy occlusion and so on in real-world.Correlation Filter(CF)based trackers have gained wide attentions due to their attractive performance both in speed and accuracy since these methods were proposed.CF based trackers have been a widely used framework for visual object tracking in real time.The main idea of CF based trackers is to train a correlation filter which can best separate the target from backgrounds through a ridge regression model.CF based trackers use the circulant samples to approximately represent the training samples so that their computational efficiency is significantly improved.In the tracking process,target sometimes shows more than one appearance,and the existing correlation filter based visual object tracking algorithms can't keep high precision to many appearances.Besides,long-term tracking requires a re-detection module.Therefore,a multiple appearances based long-term visual object tracking algorithm is proposed to improve the performance of object tracking.The main works are as follows:1)Multiple Appearance Correlation Filters(MACF)based visual object tracking algorithm are proposed.Multiple appearance models are trained by different appearances of the target during the tracking process,so that the target can be more accurately located and tracked in the case of deformation.2)A fusion tracking algorithm based on MACF and improved color histogram is proposed.In order to use color information,method is firstly performed on nonlinearly enhancement the target probability map in the search area,and then a tracking algorithm based on the improved color histogram is proposed to reduce the impact of background clutter in the target probability map.Finally,a fusion tracking algorithm based on MACF and improved color histogram is proposed.3)Two-scale search window mode and Kalman filter based redetection target algorithm is proposed.The two-scale searching window mode switches automatically to balance both accuracy and speed in the case of fast motion.A Kalman filter is used to effectively solve the case of heavy occlusion.To validate the accuracy and robustness of the proposed method,we firstly evaluate the proposed method on a large benchmark dataset OTB-2015 that contains 100 videos,and then we compare the proposed method with 9 state-of-the-art methods.Extensive experimental results show that the proposed method can handle the cases such as significant deformation,heavy occlusion,fast motion and background clutter.The proposed method performs superiorly against other methods in terms of the distance precision and overlap precision and obtaining 87.9%and 80.1%,respectively.In addition,the proposed method achieves the real-time requirements of engineering applications.