Perspective-Aware Pedestrian Detection Using Geometric Constraints in Faster R-CNN Frameworks

다국어 초록 (Multilingual Abstract)

Pedestrian detection is vital in many applications like auto-driving especially for small pedestrians in long distance. In this work, we have investigated some of the issues related to the use of Faster RCNN for the purpose of detecting pedestrian. We argue that most of the inaccuracies observed when using this model are mainly due to two reasons: (i) Excessive receptive field of feature maps for extracting precise feature of small instances of pedestrian, and (ii) lack of features to differentiate pedestrians of different spatial scales. To address the above problem, we propose a dual-branch pedestrian detection method that takes into account perspective projection. We divide images into large-scale target images and small-scale target images according to the projection method. We design a dual-branch network structure to detect targets of the two scales (large and small) respectively, and finally fuse the detection results of the two scales through the NMS (Non-Maximum Suppression) method. The method is tested on challenging pedestrian detection datasets named Caltech and KITTI and compared with different lately pedestrian detection methods. Experimental results of our proposed method on different datasets proves that our proposed method can achieve SOTA pedestrian detection results on large-scale benchmark datasets.