A framework of tracking by multi-trackers with multi-features in a hybrid cascade way

Abstract

In this paper, a general framework based on the variant Bayesian filter is proposed for the fusion of multi-trackers along with multi-features. Comparing with other fusion based trackers, this framework tries to combine feature-level fusion methods with decision-level fusion methods and obtain the tracking result in a hybrid cascade approach. Firstly, tracking results of multi-trackers are treated as prior knowledge to provide candidate samples, which could reduce the number of samples and avoid the assumption of the Gaussian distribution. Secondly, two novel candidate selection strategies, weighted voting strategy and PageRank based strategy, are proposed and run parallel to measure the similarities among different trackers under different features and achieve the fusion through the proposed correlation matrix pool and decision vectors. Unlike traditional template matching methods, these two strategies consider both the inter similarities (the similarities between template and candidates) and intra similarities (the similarities among candidates). Finally, a novel decision and update strategy with tracklet prediction-comparison is proposed in a cascaded way to cope with the consistency problem of the proposed candidate selection strategies and measure the usability of templates. Based on the proposed framework, a tracker named as MTiB is designed and tested on widely used benchmarks. The experimental results demonstrate that our proposed framework is feasible.