Multi-bandwidth Fusion For Real-time Object Tracking Algorithm

Object tracking is an important part in the field of computer vision.It is to estimate the trajectory in the subsequent video frames when given the initial information of the target.Visual object tracking has a multitude of application scenarios such as human-computer interaction and intelligent video surveillance.So it attractes much attention both at home and abroad.Traditional tracking methods in correlation filter framework are fast in speed.However,handcrafted features can not capture the semantic information effectively,making it difficult to deal with complex appearance changes and being less robust to situations when faced with severe occlusion and deformation.Convolutional features can make up for those shortcomings,but the speed is a thorny problem.Therefore,it is necessary to study further to find a trade-off between the performance and speed.Consequently,we adopt the Hierachical Convolutional Features for Visual Tracking(HCFT)as our baseline.And we propose to make further improvements in terms of feature extration,correlation filter's training,evaluation machanism and model updating.The main contributions are as follows:1.An Average Feature Energy Ration method is proposed to adaptively reduce the dimensions of convolution channels so as to improve the accuracy and speed.We select the layers Pool4 and Conv5-3 for feature extraction.Then according to the ratio between the features of search and target areas to adaptively prune the convolution channels.It is used to mitigate the inferences of redundant features on tracking task.2.An adaptive fusion strategy which utilizes multi-Gaussian filters to predict positions is proposed to accurately predict the target position.Considering the differences among training samples of different video sequences,different Gaussian bandwidth factors of training examples are applied to train multiple correlation filters separately.We adaptively fuse all the predicted results to improve the accuracy.3.Utilize the Confidence Feedback Adjustment method to adjust the learning rate and search area in time.We use the ratios of average peak-to-correlation energy(APCE)and maximum responses between two adjacent frames to feedback the tracking states.Then whether the occlusion occurs or not can be judged.According to these data,we can adjust the learning rate of the tracker in time.In case of serious occlusion or target missing,search area will be readjusted for relocation.4.We estimate all the algorithms on OTB2013 and OTB2015 standard benchmarks.The results show that all these methods can effectively improve the accuracy and speed of the tracker and gain good robustness when faced with occlusion,deformation and similar background clutters.