Robust Monocular Distance Estimation in Degraded Visual Conditions Via a Wide-and-Deep Fusion Framework

Robust visual perception remains a central challenge for autonomous systems operating under degraded environmental conditions such as fog, low light, and artificial glare. In this study, we investigate multiple strategies to enhance the resilience of vision-based models in such scenarios, using monocular distance estimation as a representative task due to its sensitivity to visual quality and availability of ground-truth metrics. A Convolutional Neural Network (CNN) is first trained on clean visual data using a single-camera robotic platform, and its direct inference results serve as the baseline. We then explore three enhancement strategies: (1) image preprocessing with a transformer-based large model, (2) model fine-tuning on low-visibility data, and (3) a novel fusion framework that integrates temporal motion history with image features. Inspired by wide-and-deep learning structures, the proposed fusion model combines shallow inputs (e.g., recent speed and displacement) with deep CNN-based image embeddings for improved robustness. Experimental results across various lighting environments demonstrate that while multiple enhancement strategies improve estimation accuracy, the fusion model achieves the most consistent and efficient performance. The findings provide insight into designing more reliable vision systems for low-cost robotics and general perception tasks under real-world variability.