Object tracking based on supervised saliency detection
Yao Zhu,Qihai Zhu,Xiaojiao Mao,Yubin Yang
DOI: https://doi.org/10.13232/j.cnki.jnju.2017.04.018
2017-01-01
Abstract:In this work,we focus on short-term single object tracking,which is the most general type of tracking problems.Numerous significant trackers have been proposed over the past few decades.As we can see from these trackers,the methods that adopt the mid-level representation have shown their superiority over other approaches in dealing with challenging factors like partial occlusion.However,most of their representation model lack the spatial information,which usually leads to poor robustness in object tracking.As a popular middle level representation,superpixels are semantically meaningful and much more homogeneous than randomly selected square patches.In this work,we construct a graph based on superpixels to introduce spatial information.A graph-based saliency detection model,which uses manifold ranking to compute the saliency scores for each superpixel for the first stage,is combined with a discriminative model,which trains a classifier to classify the candidate superpixels as target or non-target.The saliency scores and the classification result adjusted by the spatial information are then used to select the seeds for random walk as prior knowledge,which makes the result of saliency detection more relevant to the target object.Combining the classification result and the saliency scores computed by the second stage random walk,a confidence map is achieved,based on which candidates are fast ranked utilizing integral graph method.The top-ranked candidate is regarded as the target object.In order to evaluate our approach,we compare our results with other trackers along with some analysis.The experimental results on visual tracking benchmark dataset demonstrate that our approach is effective for fast motion,partial occlusion and background clutter in tracking,thus realizes desirable and robust tracking performance under complex conditions.