Multifactorial approach for robust real time object tracking

The paper presents a novel multi-factorial approach for robust real-time object tracking. The target object is modeled using joint features of color (Intensity) histogram bins, texture, shape. In subsequent frames of a video, target localization is done by generating a confidence-map (a binary image) which discriminates foreground and background using K-means clustering algorithm. Random samples (sample objects) around the previous target position are selected and modeled using joint features of color (Intensity) histogram bins, texture, shape based on confidence map. A decision matrix is generated using a set of similarity features between target and the current sample. A confidence measure of each particle is estimated using multi-factorial approach which involves a fuzzy mechanism based on weighted values of different similarity measures. Average weighted position of these samples gives position of the object in next frame of a video. The proposed technique gives better results as compared to existing mean shift and template matching based tracker under gray scale videos and is invariant to scale, translation and rotation. Results show that the proposed method improves localization error approximately by 7% over existing traditional mean shift algorithm for a gray scale video.