In highway tunnels in mountainous areas, there is insufficient illumination intensity in a closed environment at night, and after imaging, the average pixel illumination intensity is low. The data information obtained by a single sensor is usually limited. Multiple sensors improve the image fusion performance at the tunnel mouth with low illumination. Infrared sensors use the thermal radiation generated by the object to achieve automatic detection and capture the object under the condition of low illumination; visible images provide rich background information. The image information of infrared and visible light and the electromagnetic spectrum is fused to obtain enhanced and more comprehensive scene information. Image processing in a low-illumination environment has always been a hot issue in academic research. This paper used Convolution Sparse Representation (CSR), Spectral Edge (SE) and local energy features for image fusion. An intelligent sensing method for spatial information of highway tunnels under static and dynamic light environments is proposed. The denoising and fusion are processed simultaneously to avoid the loss of visible and near-infrared information during fusion processing. Bilateral filtering and light component are used for adaptive image enhancement of low-illuminance infrared and visible light source images at the tunnel mouth. Gamma correction is used to correct the illumination component to avoid distortion during image enhancement. In order to improve the visual information presented by visible tunnel light, infrared and original visible image are fused to enhance the dark details of infrared pixels. In order to further improve the feedback of multiple information in the image, the non-subsampled contour is used to decompose the preprocessed image in multi-scale and multi-direction. The non-subsampled pyramid and non-subsampled directional filter are the main components of the non-subsampled contour wave. The k-layer decomposition of the preprocessed source image, k+1 subband image with the same size can be obtained. The algorithm uses bilateral filters to decompose a single low-frequency subgraph decomposed by k layers into low-frequency basic components and detail feature components, respectively, for visible image and near-infrared images. The former is fused by local energy features, while the detail feature components are fused by convolution sparse representation strategy. The weighted local energy preserves structured information. Since simple weighting often leads to fading of infrared targets, the local feature energy ratio is used to measure the details extracted to maintain the brightness of fusion targets. A new activity measurement method and spectral edge processing were constructed at the high-frequency coefficients according to the underlying visual features; edge information is injected into the multi-source image to extract high-frequency information. Finally, the fusion coefficients were reconstructed to obtain the fused image. Four groups of visible and infrared source images captured by simulating the driver's line of sight were fused and compared with the algorithm results. The experiments were compared and analyzed from subjective evaluation and objective evaluation. Experimental results show that the CSR-SE-Energy algorithm overcomes the traditional "SR" and "pseudo-Gibbs" effects, makes up for the shortcomings of poor correlation between images, and saves Energy information and edge details. The fusion algorithm outperforms BF, SE, NSCT-BF, SF-Energy-Q and SR-C&L in subjective evaluation. The subjective visual effect has high contrast and good identification, the whole image scene can be highlighted, and the running time can be shortened. In objective evaluation, the highest MI value was 7.596 2, the highest IE value was 7.764 2, and the highest standard deviation value was 82.194 1. Compared with BF, SE, NSCT-BF, SF-energy-Q and SR-C&L algorithms. This method has significant reference significance in reducing noise, equalizing illumination and restoring details. When processing the image at the entrance and exit of the low illumination tunnel, the operation time is reduced by 0.023 2 s at most, reducing the overall operation cost and improving the image's robustness and visual clarity.