Multispectral pedestrian detection based on UNet and attention mechanism
Jiaren Guo, Z. Huang, Yanyun Tao · 2024
Pedestrian detection is commonly studied using visible light (RGB) images. However, in environments with poor lighting or occlusions, the performance of object detection using RGB images tends to be subpar. In contrast, infrared (IR) images yield better results in such special circumstances. Therefore, multispectral image fusion for pedestrian detection is increasingly being adopted. While this feature interaction and fusion approach enhances features, it can also introduce background noise, preventing the model from focusing on pedestrians. To address these issues, we propose a lightweight network with Unet and attention mechanism (UANet). UANet enhances the accuracy of pedestrian detection under varying illuminance by combining the strengths of RGB-IR image pairs. In UANet, a UNet-based module is designed for extracting and reusing multi-scale features from RGB images. To further extract high-level features, we incorporate a fully connected module into the UNet-based module. For infrared images, we employ high-frequency attention and average attention to extract target features from infrared images and fuse them with RGB features. In experiments, we validate the effectiveness of UANet in improving pedestrian detection performance. We conducted tests on the KAIST dataset, yielding a result of 95.23% mAP and 7.65% MR.