Multispectral Pedestrian Detection in Low-Light Conditions: Infrared, Visible, and Fusion-Based Approaches for CCTV Applications
Session
Mechatronics, System Engineering and Robotics
Description
Reliable pedestrian detection under low-light conditions remains a big challenge for intelligent surveillance and autonomous monitoring systems. Visible (VIS) or RGB sensor fail in darkness due to limited illuminations, while in the other hand infrared (IR) camera lack fine visual texture. This paper presents an adaptive multispectral pedestrian detection framework to address these limitations that combines the strengths of VIS and IR modalities through Weighted Boxes Fusion (WBF) applied to separately trained YOLOv8 models. Three variants of YOLOv8 the nano, small and medium models were trained independently on the LLVIP dataset. During inference, their outputs were merged using WBF with adaptive modality weighting determined by the mean brightness of each image. The models and fusions were evaluated using Precision, Recall, mAP@0.50 and mAP@0.50:0.95 as performance metrics. Results demonstrate that the adaptive fusion outperforms both single-modality detectors achieving up to 0.97 precision and 0.914 recall. Also, the correlation between image brightness and detection confidence is negligibly negative (−0.03 ≤ r ≤ −0.06), confirming that the adaptive weighting successfully neutralizes the influence of illumination on detection reliability. Providing a lightweight and scalable solution for real-time CCTV and smart-city surveillance application these findings validate the effectiveness of brightness-guided late fusion in achieving illumination-invariant multispectral pedestrian detection.
Keywords:
Multispectral pedestrian detection, Adaptive fusion, YOLOv8, Weighted Boxes Fusion, Low-light surveillance
Proceedings Editor
Edmond Hajrizi
ISBN
978-9951-982-41-2
Location
UBT Lipjan, Kosovo
Start Date
25-10-2025 9:00 AM
End Date
26-10-2025 6:00 PM
DOI
10.33107/ubt-ic.2025.290
Recommended Citation
Rexhaj, Ylli; Rexhepi, Redon; Jetullahu, Arxhend; and Rafuna, Ideal, "Multispectral Pedestrian Detection in Low-Light Conditions: Infrared, Visible, and Fusion-Based Approaches for CCTV Applications" (2025). UBT International Conference. 7.
https://knowledgecenter.ubt-uni.net/conference/2025UBTIC/MSER/7
Multispectral Pedestrian Detection in Low-Light Conditions: Infrared, Visible, and Fusion-Based Approaches for CCTV Applications
UBT Lipjan, Kosovo
Reliable pedestrian detection under low-light conditions remains a big challenge for intelligent surveillance and autonomous monitoring systems. Visible (VIS) or RGB sensor fail in darkness due to limited illuminations, while in the other hand infrared (IR) camera lack fine visual texture. This paper presents an adaptive multispectral pedestrian detection framework to address these limitations that combines the strengths of VIS and IR modalities through Weighted Boxes Fusion (WBF) applied to separately trained YOLOv8 models. Three variants of YOLOv8 the nano, small and medium models were trained independently on the LLVIP dataset. During inference, their outputs were merged using WBF with adaptive modality weighting determined by the mean brightness of each image. The models and fusions were evaluated using Precision, Recall, mAP@0.50 and mAP@0.50:0.95 as performance metrics. Results demonstrate that the adaptive fusion outperforms both single-modality detectors achieving up to 0.97 precision and 0.914 recall. Also, the correlation between image brightness and detection confidence is negligibly negative (−0.03 ≤ r ≤ −0.06), confirming that the adaptive weighting successfully neutralizes the influence of illumination on detection reliability. Providing a lightweight and scalable solution for real-time CCTV and smart-city surveillance application these findings validate the effectiveness of brightness-guided late fusion in achieving illumination-invariant multispectral pedestrian detection.