Pedestrian Detection Using an Extended Fast RCNN based on a Secure Margin in RoI Feature Maps
Mahmoud Saeidi, Ali Ahmadi · 2018
Pedestrian Detection based on Deep Convolutional Neural Network (DCNN) has recently gained a great deal of attention. Most of the proposed CNN based methods train networks employing either of the well-known Region-based CNN (RCNN) or Fast Region-based CNN (FRCNN) approaches. In this paper, we present a novel method to train Deep CNN. This method is based on an extended and improved FRCNN for pedestrian detection. It performs both classification and bounding-box regression more accurately. The proposed approach takes the advantage of a Secure Margin in Region of Interest (SM-RoI) to create multi-RoIs. Then based on some criteria, it chooses one of the RoIs with the highest score. The bounding-box extracted from the proposed FRCNN-SM approach is more effective than that of FRCNN approach in fitting and covering pedestrian. Evaluated on Caltech dataset, our proposed approach detects pedestrian more accurately than RCNN and FRCNN approaches.