Mixed local channel attention-based YOLOv8 for pedestrian detection in infrared images

In nighttime pedestrian detection tasks, objects in dark areas can be more easily detected with infrared images. However, the detailed features and contours of object are usually blurry in infrared images. Additionally, due to interference from heat sources, objects in the background with similar infrared radiation as pedestrians may overlap with them, causing confusion of detection models. To address these aforementioned issues, we propose a Mixed Local Channel Attention based YOLOv8 model (YOLOv8-MLCA) in this paper to detect pedestrians in infrared images. Firstly, we propose a Mixed Local Channel Attention (MLCA) module in YOLO's feature extraction backbone network. MLCA combines local and global information from the channel and space dimensions to enhance pedestrian details and contour features. To further address the problem of blurred pedestrian boundaries in infrared images, we propose a Minimum Points Distance based MPDLoss for bounding box regression during model training. We conducted comparative experiments on the LLVIP dataset. Among 5 baseline models in the experiment, the proposed YOLOv8-MLCA achieved the highest accuracy. We also conducted comprehensive ablation analysis to explore the performance of MLCA and MPDLoss. The experimental results validated the effectiveness of the proposed improvements.