ETD

• Feature-wise Linear Modulation (FiLM)
• Thermal Image Analysis
• Thermal Object Detection
• Transformer-based Architectures
• Vision Transformers
• Weather Conditioning
• Weather-aware Deep Learning

Thermal object detection is widely used in civilian safety, industrial inspection, and environmental monitoring, particularly in scenarios where visibility is degraded or illumination conditions are unreliable. Despite recent advances in deep learning–based detection, most existing approaches implicitly assume stationary scene statistics. In practice, however, thermal imagery is strongly influenced by environmental variability such as seasonal transitions, humidity changes, and thermal crossover effects, which can alter contrast characteristics and lead to unstable predictions.

This thesis introduces an environmentally conditioned thermal object detection framework that explicitly incorporates weather context into the detection pipeline. Building on the DETR architecture, the proposed method integrates temperature and humidity metadata through a lightweight Feature-wise Linear Modulation (FiLM) module. Rather than introducing additional sensing modalities, the model adapts its internal feature representations to environmental conditions while maintaining an end-to-end transformer based detection structure.

The resulting architecture preserves the simplicity of DETR while improving robustness under diverse acquisition scenarios.Experiments performed on the Long-Term Thermal Dataset (LTD/LTDv2) indicate a steady decrease in false detections during fog and seasonal drift circumstances, alongside more stable predictive performance over time.

The findings indicate that explicitly including environmental context can improve the operational reliability of thermal vision systems in practical civilian applications, without adding to sensor complexity or processing demands.