In view of the poor efficiency of existing pedestrian detection algorithms based on single-spectral images in all-weather environments,a pedestrian detection algorithm based on progressive multispectral images was proposed.A differential position attention enhancement module was designed,which was embedded in the dual-stream CSP-Darknet53 feature backbone extrac-tion network at different levels,to incrementally enhance global complementary information between different spectra.The cross-modal complementary information fusion strategies were used to improve existing fusion methods,global features were used to guide multi-scale feature fusion for more robust pedestrian detection.Experimental results show that the mAP50 value on the LLVIP dataset is 97.2％,and the mAP50 value on the FLIR dataset is 84.6％,indicating good detection performance.
维普
万方数据