Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Local feature based person detection and tracking beyond the visible spectrum

: Jüngling, Kai; Arens, Michael


Hammoud, R.:
Machine vision beyond visible spectrum
Berlin: Springer, 2011 (Augmented vision and reality 1)
ISBN: 978-3-642-11567-7
ISBN: 978-3-642-11568-4
Aufsatz in Buch
Fraunhofer IOSB ()
person detection; person tracking; visual surveillance; SURF

One challenging field in computer vision is the automatic detection and tracking of objects in image sequences. Promising performance of local features and local feature based object detection approaches in the visible spectrum encourage the application of the same principles to data beyond the visible spectrum. Since these dedicated object detectors neither make assumptions on a static background nor a stationary camera, it is reasonable to use these object detectors as a basis for tracking tasks as well. In this work, we address the two tasks of object detection and tracking and introduce an integrated approach to both challenges that combines bottom-up tracking-by- detection techniques with top-down model based strategies on the level of local features. By this combination of detection and tracking in a single framework, we achieve (i) automatic identity preservation in tracking, (ii) a stabilization of object detection, (iii) a reduction of false alarms by automatic verification of tracking results in every step and (iv) tracking through short term occlusions without additional treatment of these situations. Since our tracking approach is solely based on local features it works independently of underlying video-data specifics like color information - making it applicable to both, visible and infrared data. Since the object detector is trainable and the tracking methodology does not make any assumptions on object class specifics, the overall approach is general applicable for any object class. We apply our approach to the task of person detection and tracking in infrared image sequences. For this case we show that our local feature based approach inherently allows for object component classification, i. e., body part detection. To show the usability of our approach, we evaluate the performance of both, person detection and tracking in different real world scenarios, including urban scenarios where the camera is mounted on a moving vehicle.