Enhancing Night-Time Driving Safety through Deep Learning-Based Semantic Segmentation of Thermal Images

Abstract

Night-time driving presents considerable difficulties brought on by decreased visibility and illumination, elevating the possibility of mishaps. Thermal imaging technology offers a promising solution by capturing thermal radiation emitted by objects, independent of ambient lighting conditions. In this paper We suggest a unique method for the semantic division of thermal images acquired during scenarios involving driving at night using deep learning techniques. Our method, titled the "Multi-Modal Semantic Segmentation Algorithm for Night-Time Scene Understanding," leverages convolutional neural networks (CNNs) to accurately classify pixels in thermal images into meaningful categories such as roads, vehicles, pedestrians, and obstacles. We employ an encoder-decoder architecture, transfer learning, and tailored data augmentation strategies to improve generality along with accuracy of segmentation capability. Tests conducted using publically accessible datasets, including the KAIST dataset, demonstrate the effectiveness of our approach in accurately segmenting thermal images. Performance metrics such as pixel-level accuracy (99%), mean intersection over union (mIoU) (95%), overall precision (95.75%), overall recall (96.25%), overall F1 score (95.75%), accuracy (98%), and generalization accuracy (97%) are included in the detailed findings section of the paper. These values provide quantitative measures of the effectiveness of the proposed approach, showcasing its superiority over existing techniques in terms of accuracy and computational efficiency. Our research contributes to improving night-time driving safety and advancing autonomous vehicle technology.