Comparative Analysis of State-Of-The-Art Object Detection Models Under Varying Lighting and Camera Angles

Yalçın Y., Gölcük B., Urhan O.

15th INTERNATIONAL MARMARA SCIENCES CONGRESS, Kocaeli, Türkiye, 28 - 29 Kasım 2025, ss.138-145, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Basıldığı Şehir: Kocaeli
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.138-145
  • Kocaeli Üniversitesi Adresli: Evet

Özet

The performance of object detection algorithms is significantly affected by environmental conditions that differ from the training data. In this study, the performances of state-of-the-art deep learning-based object detection models are investigated under varying illumination levels and camera fields of view (FoV). Within the scope of the study, a unique dataset comprising 10 classes was created using lenses with viewing angles of 70°, 125°, and 180° under illumination intensities of 3, 50, and 90 lux. The detection performance of the YOLOv9, YOLOv11, and RF-DETR models was compared on the collected data. Experimental results showed that detection accuracy improved as ambient light intensity increased, with the improvement more pronounced in wide-angle shots. However, as the field of view (FoV) increased, the object dimensions in the image decreased, leading to a decline in the models’ detection precision. In the conducted tests, the highest performance was achieved under 90 lux illumination with 70° and 125° viewing angles.