Discriminative Feature Transformation for Occluded Pedestrian Detection

Despite promising performance achieved by deep convolutional neural networks for non-occluded pedestrian detection, it remains a great challenge to detect partially occluded pedestrians. Compared with non-occluded pedestrian examples, it is generally more difficult to distinguish occluded pedestrian examples from backgrounds in featue space due to the missing of occluded parts. In this paper, we propose a discriminative feature transformation which enforces feature separability of pedestrian and non-pedestrian examples to handle occlusions for pedestrian detection. Specifically, in feature space it makes pedestrian examples approach the centroid of easily classified non-occluded pedestrian examples and pushes non-pedestrian examples close to the centroid of easily classified non-pedestrian examples. Such a feature transformation partially compensates the missing contribution of occluded parts in feature space, therefore improving the performance for occluded pedestrian detection. We implement our approach in the Fast R-CNN framework by adding one transformation network branch. We validate the proposed approach on two widely used pedestrian detection datasets: Caltech and CityPersons. Experimental results show that our approach achieves promising performance for both non-occluded and occluded pedestrian detection.