A Study of Exploiting Objectness for Robust Online Object Tracking

Tracking is a fundamental problem in many computer vision applications. Despite the progress over the last decade, there still exist many challenges especially when the problem is posed in real world scenarios (e.g., cluttered background, occluded objects). Among them drifting has been widely observed to be a problem common to the class of online tracking algorithms - i.e., when challenges such as occlusion or nonlinear deformation of the object occurs, the tracker might lose the target completely in subsequent frames in an image sequence. In this work, we propose to exploit the objectness to partially alleviate the drifting problem with the class of online object tracking and verify the effectiveness of this idea by extensive experimental results. More specifically, a recently developed objectness measure was incorporated into Incremental Learning for Visual Tracking (IVT) algorithm in a principled way. We have come up with a strategy of reinitializing the training samples in the proposed approach to improve the robustness of online tracking. Experimental results show that using objectness measure does help to alleviate its drift to background for certain challenging sequences