Research On Correlation Filter-based Visual Object Tracking

Despite excellent performance which has been shown by discriminative correlation filter-based visual trackers,most of the trackers still suffer from the three defects including insufficient learning,inauthentic combination and inflexible update when it comes to multiple feature channels.In this paper,we propose IFCT(Independent Feature Channels-based Tracking)with respect to the above three defects.A specific correlation filter is learned independently in each one feature channel for tackling the defect of insufficient learning,then the final response map is generated according to the confidence of feature channels for a solution to inauthentic combination.The confidence of feature channels is decided by both the optimal spatial distribution and the optimal temporal distribution.Under this situation,correlation filters in all feature channels are allowed to be online updated in separate channels with their own learning rates.Adaptive learning rates exploited are capable of reflecting the diversity of the variant feature space of corresponding feature channels.Extensive experiments are conducted on OTB-2013 and OTB-2015 datasets.Our IFCT outperforms the baseline fDSST(fast Discriminative Scale Space Tracking)by 4.5%,5.2% and 6.1% respectively in AUC(Area-Under-the-Curve),DP(Distance Precision)and OP(Overlap Precision).At the same time,our approach shows competitive performance with comparison to other state-of-the-art trackers while running at 70 FPS.Furthermore,almost all correlation filter-based trackers suffer severely from model drift.Model drift is a difficult problem which happens during the process of visual tracking.In this paper,model drift is classified into two kinds: scend-caused model drift and algorithm-caused model drift.To tackle scene-caused model drift,correlation filters with retrogression mechanism are exploited.Model drift brings samples which are not helpful for correlation filters to discriminate the object,and this kind of samples are called impure samples in this paper.Model drift is eliminated by allowing contaminated correlation filters to retrogress to the ones that are not influenced by impure samples.As a result,correlation filters are purified by neglecting the impure part of them.By experiments conducted on both OTB-2013 and OTB-2015,the effectiveness of retrogression mechanism is proved.Our approach outperforms the baseline by 2.8%,3.2% and 3.5% respectively in AUC,DP and OP.To take care of both scene-caused model drift and algorithm-caused model drift,RCF(Retrogression of independent feature channels-based Correlation Filters)is proposed.With respect to algorithm-caused model drift,RCF is expected to improve the original algorithm to reduce the number of failure cases.As a result,IFCT is introduced to improve the correlation filter-based tracking algorithm by suppressing the defects including insufficient learning,inauthentic combination and inflexible update.To scene-caused model drift,retrogression mechanism above is introduced to eliminate the experience limiting the ability of correlation filters discriminating.Besides the above effective combination between IFCT and retrogression mechanism of correlation filters,RCF also takes confidence level of detected results and variant level of interframe feature space into consideration when it comes to online update.Our approach is experimented on both OTB-2013 and OTB-2015.Our RCF outperforms the baseline fDSST by 5.6%,6.9% and 8.1% respectively in AUC,DP and OP on OTB-2015 and shows competitive performance compared to other state-of-the-art trackers.