Analysis of the impact of rain on perception in automated vehicle applications

The reliable performance of object detection perception algorithms in automated vehicles under adverse conditions such as rain is critical for maintaining vulnerable road user safety. Visible-spectrum cameras provide a rich source of information and are cost-effective compared with other sensors; however, their performance can degrade under adverse environmental conditions. Despite the general consensus that the object detection performance in computer vision is adversely affected by rain, there is a relative lack of research investigating this relationship in detail. This study investigates the performance of object detection under rain conditions, focusing on algorithm performance and low-level object characteristics. Using the publicly available BDD100 k dataset, this study examines object detection performance across multiple deep?learning object detection architectures, analyzing error types and image characteristics under rain and no rain conditions. In addition, statistical methods were used to compare image-level metrics to determine statistical significance. The results reveal that rain is not detrimental to object detection performance, and in some cases, better performance is observed. For some models, medium-sized objects experience improved detection and classification under rain conditions, while large objects experience a slight decline in performance. The error analysis shows an increase in localization errors and a decrease in classification errors. The object-level analysis revealed statistically significant changes in the contrast-to-noise ratio, entropy, mean pixel value, pixel variance, hue, saturation, and value, with hue and saturation experiencing the most significant changes. This study highlights the need for more detailed weather labeling in datasets to fully understand the nuances of the relationship between rain and object detection.