Pedestrian Detection in Nighttime Infrared Images Based on Improved YOLOv8 Networks
Zhenyue Zhang, Boran Wang, Wensheng Sun · 2023
Nowadays, due to poor nighttime lighting conditions, common visible light pedestrian detection mechanisms exhibit low sensitivity, leading to potential safety hazards. To address this issue, this paper proposes a nighttime pedestrian detection method called YOLOv8_FastD based on infrared images. This method constructs a C2f_FastD module using deformable convolution and partial channel convolution on the basis of the YOLOv8 backbone network. This module enhances the learning of target features in complex backgrounds, reduces redundant computations, and improves detection speed. Additionally, a squeeze-and-excitation (SE) channel attention layer is appended after each C2f_FastD block in the backbone network to further enhance detection accuracy. Experimental results on the open-source infrared image dataset Kaist demonstrate that the proposed method achieves a detection accuracy of 93.4% and a detection frame rate of 78 frames per second (FPS). Compared to the YOLOv8 model, the proposed method shows an improvement of 6.8% in detection accuracy and a 81.4% increase in FPS.