Assessing the Detection Capabilities of RGB and Infrared Models for Robust Occluded and Unoccluded Pedestrian Detection

Pedestrian detection is a critical component of autonomous vehicle systems, particularly under challenging conditions where visibility may be compromised due to occlusion or low-light. This study presents a comparative analysis of RGB (visible spectrum) and infrared (IR) modalities, each employed independently to detect pedestrians under various conditions. By evaluating single-modal models on both occluded and unoccluded pedestrians, we aim to provide insights into each modality’s strengths and limitations. Through extensive training, validation, and testing phases on the FLIR dataset, the RGB model achieved mean Average Precision ([email protected]) of 0.8711, and 0.8739 during training and testing, while the IR model recorded a higher training [email protected] of 0.9020, and 0.8729 during testing. Notably, the IR model achieves higher [email protected], indicating greater robustness in scenarios with relaxed bounding box precision requirements. These findings suggest that integrating RGB and IR modalities in a complementary manner could enhance detection reliability across diverse real-world conditions. This work provides a foundation for developing more resilient and accurate pedestrian detection systems in autonomous vehicles, capable of operating effectively under both occluded and unoccluded conditions.