Robust Visual Tracking by Exploiting the Historical Tracker Snapshots

Variations of target appearances due to illumination changes, heavy occlusions and abrupt motions are the major factors for tracking failures. In this paper, we show that these failures can be effectively handled by exploiting the trajectory consistency between the current tracker and its historical trained snapshots. Here, we propose a Scale-adaptive Multi-Expert (SME) tracker, which combines the current tracker and its historical trained snapshots to construct a multi-expert ensemble. The best expert in the ensemble is then selected according to the accumulated trajectory consistency criteria. The base tracker estimates the translation accurately with regression based correlation filter, and an effective scale adaptive scheme is introduced to handle scale changes on-the-fly. SME is extensively evaluated on the 51 sequences tracking benchmark and VOT2015 dataset. The experimental results demonstrate the excellent performance of the proposed approach against state-of-the-art methods with real-time speed.