Best-match rectangle adjustment algorithm for persistent and precise correlation tracking

Correlation tracking is performed by matching a template of the object with every candidate region in a search window and locating the best-match region. The template is initialized with the target manually, when it appears in the video. The human operator is usually unable to extract a good template of the target within a limited time period, while the target is moving and manoeuvring in the streaming video. Thus, the extracted template is usually larger or smaller than the object, or the object is deviated from its centre. Furthermore, if the object is manoeuvring in the video during a tracking session, it drifts away from the centre of the template with time, even when the template was initialized correctly. The incorrect template initialization and the template-drift severely deteriorate the performance of a correlation tracker. We solve the two problems by analyzing the temporally smoothed edge-enhanced template and resizing/relocating the best-match rectangle (BMR) so that the object can remain exactly at its centre. The adjusted BMR is then exploited to generate a good template for the next frame to have a persistent and precise tracking session. We present the real-world results that prove the efficacy of the proposed algorithm.