Traffic Object Detection Across Diverse Environmental Conditions: Comparative Performance Evaluation of Computer-Vision Models

This paper introduces a novel evaluation framework for systematically assessing the performance of computer-vision models in traffic-surveillance and autonomous-vehicle applications under varying environmental conditions. The framework integrates synthetic-data generation with structured experiments to analyze model performance across different weather and lighting scenarios, addressing the lack of controlled datasets for true apple-to-apple comparisons. The results obtained based on synthetic and real-world video data reveal key trade-offs between detection accuracy, inference time, and environmental resilience across different model categories. One-stage models, such as YOLO, exhibit high inference speed but suffer significant performance degradation in adverse weather, whereas RetinaNet and SSD demonstrate greater robustness. Two-stage models, particularly Faster R-CNN, provide higher accuracy and stability but require longer inference times. Diffusion-based models fail in surveillance-camera views, highlighting training limitations. Also, illumination changes affect detection performance less than rain and snow, likely because of better representation of lighting variations in training datasets. This work is among the first to systematically quantify performance degradation under controlled environmental conditions, providing valuable insights for real-world deployments in intelligent transportation systems.