Enhancing Object Detection Robustness In Adverse Weather Conditions

Abstract

Fundamental applications in autonomous cars, robots, and surveillance systems include object detection for safety reasons such as recognizing vehicles, pedestrians, trees, street poles, and buildings. However, one of the major challenges it faces from adverse weather conditions like rain, haze, and fog through deterioration of accuracy and an increase in safety risks such as traffic accidents. This paper proposes an enhanced object detection system to improve its robustness and accuracy in challenging environments. The proposed approach effectively addresses the problem of poor image quality due to adverse weather conditions by using unsupervised domain adaptation and including a Quasi-Translation Network and Feature Calibration Network in the YOLO(You Only Look Once) framework. It attempts to leverage the success made in CNNs(Convolutional Neural Networks) and mass amounts of datasets in order to adapt YOLO-series detectors in the best possible way, achieving object detection under challenging environmental conditions with precision and speed. Experimental results demonstrate that the solution proposed outperforms existing models in terms of adaptability and accuracy, providing a scalable and reliable solution for autonomous systems to enhance overall safety and operational efficiency in real-world environments.