Machine Learning Based Bounding Box Regression for Improved Pedestrian Detection

The most of the studies on pedestrian and passenger detection focus on end-to-end learning by considering either improvement of features to be used or the enhancement of the detectors. One of the important steps of these systems is non-maximum suppression (NMS), which aims reducing proposed bounding boxes that supposed to belong the same target through a greedy regional search and clustering. In order to improve the performance of NMS, recent approaches consider using only bounding boxes and their scores. By following this path with a novel approach, in this study, a machine learning based bounding box regression approach is proposed. During the training phase, proposed system uses position, size and confidence scores of bounding boxes as features and the same information of the corresponding ground truth (except score) as the desired output. By this way, a pattern between initially generated bounding boxes and the ground truth is revealed. Several tests and experiments have been performed and the results show that the developed system can be particularly effective when correct decisions are needed with low overlapping ratios (such as applications with strong occlusion) without increasing false positives.