Context Constraint & Pattern Memory for Long-term Correlation Tracking

Abstract This paper proposes a novel correlation tracking framework with context constraint and pattern memory (CCPM) to address scale variation, heavy occlusion and out-of-view in long-term tracking tasks. Taking the spatial structure into account, we learn multi-channel correlation filters by training the background-aware samples constrained by Gaussian function and utilize the scaling pool technique to estimate the target scale variation. Then a novel online recovery mechanism, including the weighted fern classifier and the pattern-memorized template matcher, is applied to strengthen the ability to re-capture the target in case of tracking failure. The spatio-temporal weights used in the fern classifier can effectively suppress the significant position drift caused by the distortion of the target, whereas the pattern-memorized templates compressed by Gaussian mixture model (GMM) can reduce capacity usage and computational cost. Finally, we focus on the model corruption problem, the average peak-to-correlation energy (APCE) is exploited to identify the reliable parts of the tracking trajectory, and we update the model adaptively in term of the feedback from high-confidence tracking results. Extensive experimental results demonstrate that the proposed algorithm performs favorably in OTB2013 data set against several state-of-the-art methods.