Improving Person Detection in Videos by Automatic Scene Adaptation
Publikation aus Digital
Mörzinger R., Thaler M.
Proceedings of VISAPP2010, 2010
The task of object detection in videos can be improved by taking advantage of the continuity in the data stream, e.g. by object tracking. If tracking is not possible due to missing motion features, low frame rate, severe occlusions or rapid appearance changes, then a detector is typically applied in each frame of the video separately. In this case the run-time performance is impaired by exhaustively searching each frame at numerous locations and multiple scales. However, it is still possible to significantly improve the detectors performance if a static camera and a single planar ground plane can be assumed, which is the case in many surveillance scenarios. Our work addresses this issue by automatically adapting a detector to the specific yet unknown planar scene. In particular, during the adaptation phase robust statistics about few detections are used for estimating the appropriate scales of the detection windows at each location. Experiments with an existing person detector based on histograms of oriented gradients show that the scene adaptation leads to an improvement of both computational performance and detection accuracy. For scene specific person detection, changes to the implementation of the existing detector were made. The code is available for download. Results on benchmark datasets (9 videos from i-LIDS and PETS) demonstrate the applicability of our approach.