Fast visual people tracking using a feature-based people detector

This article presents a new technique to visual people tracking that combines feature-based object detection with feature tracking. In comparison to tracking by detection techniques which are frequently employed by feature-based approaches, the explicit tracking of detected features achieves a speed up of the overall tracking process and allows to adapt the chosen features on-line; this way, the tracker can specialize itself to tracking a specific object and gain an improved robustness in multi-object scenarios. Our approach uses a fast version of an implicit shape model detector (ISM) trained on people to initially find a person. The tracking process itself is then based on tracking the features found by the detector. Features are tracked by carrying out a correspondence search from frame to frame in combination with a voting scheme for determining a person's center position in the image. The approach does not require a motion model for the objects and is robust against bumpy or jerky motions of the camera. To further improve robustness, the feature set is continuously updated in order to adapt to changes in appearance. Experiments show that the proposed system is able to track multiple people. It also can re-identify persons after periods of occlusion and distinguish them from each other, even if they are looking similar. The system was tested on the BoBoT dataset and on a small mobile robot following a person traversing rough grassland