Pedestrian detection at daytime and nighttime conditions based on YOLO-v5

Main Article Content

Bryan Montenegro https://orcid.org/0000-0002-4514-4089
Marco Flores https://orcid.org/0000-0001-7507-3325

Abstract

This paper presents new algorithm based on deep learning for daytime and nighttime pedestrian detection, named multispectral, focused on vehicular safety applications. The proposal is based on YOLO-v5, and consists of the construction of two subnetworks that focus on working with color (RGB) and thermal (IR) images, respectively. Then the information is merged, through a merging subnetwork that integrates RGB and IR networks to obtain a pedestrian detector. Experiments aimed at verifying the quality of the proposal were conducted using several public pedestrian databases for detecting pedestrians at daytime and nighttime. The main results according to the mAP metric, setting an IoU of 0.5 were: 96.6 \% on the INRIA database, 89.2 % on CVC09, 90.5 % on LSIFIR, 56 % on FLIR-ADAS, 79.8 % on CVC14, 72.3 % on Nightowls and 53.3 % on KAIST.