Regresar

Pedestrian detection at night based on faster R-CNN and far infrared images

Abstract:

This paper presents a pedestrian detection system focused on night time conditions for vehicular safety applications. For this purpose we analyze the performance of the recent deep learning detector Faster R-CNN [1] with infrared images for detecting pedestrians at night. We discovered that Faster R-CNN has drawbacks when detecting pedestrians that are far away. For this reason, we present a new Faster R-CNN architecture focused on multi scale detection, through two Region Proposal Networks RPNCD and RPNLD. Our architecture has been compared with the best models such as VGG-16 and Resnet 101. The experimental results have been development on the CVC-09 dataset [2]. These show an improvement when detecting far away pedestrians, with a 37.73% miss rate on 10-2 FPPI and the mAP is 85%.