Nighttime agglomerate fog event detection considering car light glare based on video

Agglomerate fog event poses more serious threat than normal foggy weather to expressway traffic safety, due to its localized nature and suddenly uneven formation. However, vision-based fog detection methods typically estimate visibility for individual images and ignore the difference in the characteristics of even and uneven fog, lacking use of temporal information to differentiate between normal foggy weather and agglomerate fog events. Meanwhile, detection of fog at night faces strong interference from car lights that is always overlooked. This study proposes a nighttime agglomerate fog event detection (AFED) method for videos, taking into account car light interference. Depth disparity feature is constructed based on the information entropy of depth estimation result. In order to build a metric for uneven characteristics in the field of view, we creatively introduce the Moran’s index to establish uneven feature, generating two-dimensional feature time series for each video. By extracting interpretable features from the two-dimensional feature time series after removing car light interference frames, a classification model based on extreme gradient boosting (XGBoost) is built to differentiate agglomerate fog, normal fog, and no fog videos. Experiments are carried out utilizing real monitoring data from roadside surveillance cameras to validate the effectiveness of features and model. Furthermore, a fog event detection dataset containing over 1 500 videos is established, making up data scarcity for vision-based agglomerate fog event detection and providing support for future research.