State-of-the-art machine learning models of traffic monitoring focus on single-modal methods using only video, video-image or audio resources. However, audio analysis can only provide general classification of traffic conditions and is not reliable during dense traffic volumes and when certain events, such as helicopters passing by; video-image analysis faces challenges of poor illumination situations and inclement weather conditions. Besides the limitations of these methods, due to technical issues, one of the modalities could be missing or not informative for a specific period of time in a realistic scenario.

Thus, when the cameras or the microphones on the roads meet technical issues or bad environmental conditions, how can we estimate traffic conditions?