Automatic Detection of Object of Interest and Tracking in Active Video

Abstract

We propose a novel method for automatic detection and tracking of Object of Interest (OOI) from actively acquired videos by non-calibrated cameras. The proposed approach benefits from the object-centered property of Active Video and facilitates self-initialization in tracking. We first use a color-saliency weighted Probability-of-Boundary (cPoB) map for keypoint filtering and salient region detection. Successive Classification and Refinement (SCR) is used for tracking between two consecutive frames. A strong classifier trained on-the-fly by AdaBoost is utilized for keypoint classification and subsequent Linear Programming solves a maximum similarity problem to reject outliers. Experiments demonstrate the importance of Active Video during the data collection phase and confirm that our new approach can automatically detect and reliably track OOI in videos.