An Improved Method for Infrared Vehicle and Pedestrian Detection Based on YOLOv5s

Abstract

The urban road scene at night is complex, and the imaging effect based on visible light at night is insufficient. Aiming at the problems of lack of color information, texture detail, and low spatial resolution of infrared scenes, we propose an improved infrared detection model based on YOLOv5s. The experimental data applies infrared images that are less affected by light conditions. In order to employ the features of the images as much as possible, this experiment combines the features of the DenseNet network and YOLOv5s, replacing the Bottleneck module in the C3 module of YOLOv5s with a custom Denseblock module to derive a custom C3-Dense module. In this paper, the C3-Dense module is a substitute for the C3 partly in the backbone network to improve the feature extraction capability of the network; in order to improve the detection capability of small targets, the SE-Net module is added to the backbone network. The simplified BiFPN structure replaces the original PAN structure, which enhances the ability of the network to extract features for different resolutions. Train and test using the FLIR infrared pedestrian and vehicle dataset. The results of our experiment show that the improved model has improved recall, precision, and mean Average Precision (mAP) for pedestrian and vehicle target detection in infrared images. The recall and precision of pedestrian detection increase by 4.24% and 5.01%; for vehicle detection, that is 1.51% and 3.23%, and the mAP is increased by 3.49%.