Object Tracking Algorithm Based on a New Robust Feature

The classical mean-shift tracking algorithm is based on histogram of colors, which is vulnerable to light change. In order to overcome the drawback, we presents a new metric used for tracking. Firstly, we compute the curvature property of an image and choose scale through maximizing the second derivative in horizontal direction of points in the inner elliptical region on the target. Then we compute the second derivative of all the points in the image within selected scale and form a weight image, which reduces the weights of the objects with size that vary from the tracking target´s and protrudes the tracking target. Finally, we track the target within mean-shift framework. Several experiments on PETS database show that: our algorithm can tackle light change, is robust to partial occlussion, and is adaptive to rotation.