Recent advances in
pedestrian detection
B. Schiele
Object detection methods have been applied to a wide range
of categories. Viewpoint changes and articulated objects pose particular challenges.
Here we will report on two new improvements of our pedestrian detection method,
building upon our earlier pedestrian detection papers (@cvpr05 and @cvpr06).
Firstly, we report on a learning scheme which allows to effectively generalize
across articulations and secondly, we extend the approach with an SVM-based
verification scheme to increase detection accuracy.
Taken together, these extensions significantly outperform previous results.
These versions of our algorithm and previous work are compared on challenging
multi-viewpoint pedestrian data, including significant overlaps between pedestrians.