Long-term multiobject tracking using alternative correlation filters

Abstract

We propose a real-time multiobject-tracking approach that is minimally affected by environmental conditionsand target appearance change. The aim of the proposed approach is to track any object in a scene, regardless of objecttype, since tracking all of the objects in a scene is critical and widely used in surveillance applications. Thus, motiondetection results are used to initialize the trackers. The proposed object-tracking approach is realized with two typesof independent correlation filters estimating location and scale. Alternative correlation filters representing differentappearances of the target are also proposed in order to increase the robustness of the approach to scene and targetchanges. Tracking sustainability is provided by putting alternative correlation filters into use when the quality of thetracking output decreases to a critical level. Motion blobs are also used to minimize object boundary drift, whichis a challenging problem, especially for long-term tracking. The proposed approach was tested on an object-trackingbenchmark dataset and outperformed most state-of-the-art methods.