We propose a novel technique based on the quantitative detection of rain intensity from images, i.e., from pictures taken in rainy conditions. The method is fully analytical and based on the fundamentals of camera optics. A rigorous statistical framing of the technique allows one to obtain the rain rate estimates in terms of expected values and associated uncertainty. We show that the method can be profitably applied to real rain events, and we obtain promising results with errors of the order of 625%. A precise quantification of the method's accuracy will require a more systematic and long-term comparison with benchmark measures. The significant step forward with respect to standard rain gauges resides in the possibility to retrieve measures at very high temporal resolution (e.g., 30 measures per minute) at a very low cost. Perspective applications include the possibility to dramatically increase the spatial density of rain observations by exporting the technique to crowdsourced pictures of rain acquired with cameras and smartphones.

Monitoring road conditions, e.g., water build-up due to intense rainfall, plays a fundamental role in ensuring road safety while increasing resilience to the effects of climate change. Distributed cameras provide an easy and affordable alternative to instrumented weather stations, enabling diffused and capillary road monitoring. Here, we propose a deep learning-based solution to automatically detect wet road events in continuous video streams acquired by road-side surveillance cameras. Our contribution is two-fold: first, we employ a convolutional Long Short-Term Memory model (convLSTM) to detect subtle changes in the road appearance, introducing a novel temporally consistent data augmentation to increase robustness to outdoor illumination conditions. Second, we present a contrastive self-supervised framework that is uniquely tailored to surveillance camera networks. The proposed technique was validated on a large-scale dataset comprising roughly 2000 full day sequences (roughly 400K video frames, of which 300K unlabelled), acquired from several road-side cameras over a span of two years. Experimental results show the effectiveness of self-supervised and semi-supervised learning, increasing the frame classification performance (measured by the Area under the ROC curve) from 0.86 to 0.92. From the standpoint of event detection, we show that incorporating temporal features through a convLSTM model both improves the detection rate of wet road events (+ 10%) and reduces false positive alarms ($$-$$ -  45%). The proposed techniques could benefit also other tasks related to weather analysis from road-side and vehicle-mounted cameras.

No abstract