Aiming at the visual adaptation challenges of autonomous vehicles caused by abrupt illumination changes at highway tunnel portals, a progressive infrared and visible image fusion method in the nonsubsampled shearlet transform (NSST) domain was proposed. The method constructed a multi-scale, multi-directional hierarchical decomposition to obtain coarse-to-fine low-frequency structural information and multi-directional high-frequency details. For low-frequency coefficients, convolutional sparse representation was enhanced by integrating a block-wise processing mechanism with alternating direction method of multipliers (ADMM) optimization, effectively improving global feature consistency. For high-frequency coefficients, spectral edge information was incorporated and text ure details were preserved through a novel activity measure. The fused image was reconstructed via inverse non-subsampled contourlet transform. Experiments on real data from the G22 Ningxia Liupan Mountain Tunnel show that, compared with traditional algorithms, the proposed method improves mutual information (MI), transfer of edge information (Q^AB/F), and natural image quality evaluator (NIQE) by up to 0.98, 0.22, and 2.2, respectively, while reducing runtime by up to 17.5%. It effectively mitigates the “block effect” and “pseudo-Gibbs effect” and enhances inter-image correlation. When fed into the YOLO11s object detection network, the fused images achieve a mAP@0.5 of 92.45%, demonstrating clear benefits for downstream perception tasks. In conclusion, the proposed approach is helpful to enhance the perception accuracy of autonomous driving systems in low-illumination, high-dynamic-range environments, thereby improving driving safety and traffic efficiency.