Low-light conditions significantly degrade the performance of RGB cameras, particularly in applications like visual odometry estimation for autonomous vehicles, where feature visibility is critical for accurate navigation. Traditional image enhancement techniques often require manual tuning and struggle in extremely dark environments. However, deep-learning-based methods, though promising, typically demand high computational resources, making them unsuitable for real-time applications. This paper presents top low-light image enhancement (LLIE) models obtained through our experimentation with designing models to improve the visibility of features crucial for odometry tasks while minimizing computational overhead. We explore design and training strategies for real-time low-light enhancement using U-Net, a CNN architecture, alongside other CNN models with residual blocks and attention mechanisms. Through experimentation with two datasets-LOL-v2 and KITTI datasets, we demonstrate that our models offer significant improvements in feature visibility without compromising real-time performance. We also report that training the model on image patches reduces Graphics processing unit (GPU) memory usage and improves image enhancement quality, though training on full images may sometimes be necessary. Compared to a high-performing baseline model, our approach-despite yielding lower image quality is better suited for real-time applications. Some use cases, such as robotic navigation, can benefit from our lightweight model, where high-resolution details are less critical, as the focus is on real-time performance and general feature visibility. Additionally, we find that using U-Net without residual blocks and attention mechanisms results in degraded image quality. Future work will focus on transfer learning our model for odometryspecific image enhancement and integrating it into autonomous localization systems to optimize computational efficiency and enhance performance.