Self-supervised and semi-supervised learning for road condition estimation from distributed road-side cameras.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Environmental Monitoring

Background:

• Road safety is critical, especially with increasing extreme weather events.
• Traditional weather monitoring is limited; distributed cameras offer a scalable solution.
• Automated detection of road conditions like water build-up is needed for climate change resilience.

Purpose of the Study:

• To develop a deep learning model for automatic detection of wet road events using surveillance camera video streams.
• To enhance the robustness of the detection model against varying outdoor illumination conditions.
• To leverage self-supervised learning for improved performance on large, unlabeled video datasets.

Main Methods:

• Utilized a convolutional Long Short-Term Memory (convLSTM) model to analyze temporal changes in road appearance.
• Implemented a novel temporally consistent data augmentation technique for improved robustness.
• Developed a contrastive self-supervised learning framework tailored for surveillance camera networks.
• Validated the approach on a large-scale dataset of approximately 2000 daily video sequences (400K frames).

Main Results:

• Self-supervised and semi-supervised learning significantly improved frame classification performance (Area Under ROC curve from 0.86 to 0.92).
• Incorporating temporal features with convLSTM enhanced wet road event detection rates by 10%.
• The convLSTM model reduced false positive alarms by 45%.

Conclusions:

• The proposed deep learning solution effectively detects wet road events using road-side cameras.
• The combination of convLSTM and self-supervised learning offers a robust and accurate method for road condition monitoring.
• This approach has potential applications in broader weather analysis using camera networks.