Nighttime pedestrian detection using crossed wavelet fusion networks

Pedestrian detection at night is crucial for applications such as intelligent transportation and public safety. However, traditional pedestrian detection methods often encounter significant challenges in low light and complex background conditions. Specifically, a single visible image is hampered by insufficient lighting, leading to a marked decline in pedestrian detection performance. To address this issue, we propose a pedestrian detection method that fuses infrared and visible images. This study has holistically designed the image fusion network, CWFusion, as well as the pedestrian detection network, nighttime pedestrian detection-YOLO (NPD-YOLO). CWFusion comprises the transform fusion network, effective image fusion transformer (EIFT), and the cross-attention module, wavelet-based attention refinement (WAR), which are employed to process high-frequency information. The input image is decomposed into a low-frequency component and a set of high-frequency components using a pooling method guided by the 2D discrete wavelet transform. The EIFT module is designed to extract deeper features from the low-frequency components, effectively retaining essential global information, whereas the WAR module enhances the detailed representation of the high-frequency components through a cross-attention mechanism. The final fusion and image reconstruction are executed using a Laplace pyramid-based image fusion strategy. NPD-YOLO incorporates a feature extraction strategy and a multi-scale detection method based on YOLOv8. Our method achieves an average accuracy of 67.47% on the NVIDIA RTX 3080. In both quantitative and qualitative experiments, our approach outperforms other image fusion-based nighttime pedestrian detection models.