Vehicle and pedestrian detection method based on Frequency domain and spatial Domain cross-modal fusion

Aiming at the problems of background interference caused by insufficient illumination of visible images, consistency of thermal radiation and ghost of infrared images, this study proposes a vehicle and pedestrian detection method DLF-YOLOv11 based on cross-modal fusion of frequency domain and spatial domain. Firstly, aiming at the background interference caused by multi-source image spatial domain feature alignment, a Multi-domain fusion Attention module based on wavelet transform (DMFFA) is designed. The image features are decoupled into high frequency components and low frequency components by discrete wavelet transform, and then the fine-grained fusion of cross-modal features is realized by means of the designed cross-modal coordinated coordinate attention mechanism. Secondly, in order to further improve the edge feature extraction ability of the model, a Lightweight Edge convolution module (LEConv) is proposed. The direction-sensitive structure enables the convolutional network to capture weak edge features to further reduce background interference. Experimental verification shows that the mAP50 index of DLF-YOLOv11 on LLVIP and FILR datasets reaches 97.7%and 80.0%, respectively. Compared with the benchmark model YOLOv11, the detection accuracy of visible images is improved by 10.7% and 15.0%, and the detection accuracy of infrared images is improved by 4.1% and 0.8%, which fully proves the effectiveness of this method in cross-modal information complementary fusion.