Thermal Image-Based Multi-Class Semantic Segmentation for Autonomous Vehicle Navigation in Restricted Environments

Technological advancements have propelled the development of environmentally friendly transportation, with autonomous vehicles (AVs) and thermal imaging playing pivotal roles in achieving sustainable urban mobility. This study explores the application of the SegNet deep learning architecture for multi-class semantic segmentation of thermal images in constrained environments. The methodology encompasses data acquisition using a thermal camera in urban settings, annotation of 3,001 thermal images across 10 object classes, and rigorous model training with a high-performance system. SegNet demonstrated robust learning capabilities, achieving a training accuracy of 96.7% and a final loss of 0.096 after 120 epochs. Testing results revealed strong performance for distinct objects like motorcycles (F1 score: 0.63) and poles (F1 score: 0.84), but challenges in segmenting complex patterns such as buildings (F1 score: 0.34) and trees (F1 score: 0.42). Visual analysis corroborated these findings, highlighting strengths in segmenting well-defined objects while addressing difficulties in handling variability and elongated structures. Despite these limitations, the study establishes SegNet's potential for thermal image segmentation in AV systems. This research contributes to the advancement of computer vision in autonomous navigation, fostering sustainable and green transportation solutions while emphasizing areas for further refinement to enhance performance in complex environments.