Quanser Interactive Lab Video Clips

Efficient traffic monitoring is crucial for improving route capacity, alleviating congestion, and decreasing the probability of accidents in smart cities and Intelligent Transportation Systems (ITS). In an urban traffic environment and ITS, the dataset is mainly collected from edge surveillance cameras, and intelligent models integrated into these devices utilize this data for various use cases, such as traffic management, safety, monitoring, etc. However, traditional machine learning models, which use object detection, tracking, or a heuristic approach, do not achieve optimal performance in complex, dynamic environments. To address these problems, we present a novel framework that integrates video-based perception and language-model reasoning to enhance intelligent traffic monitoring and safety. This system employs the Video Joint Embedding Predictive Architecture (VJEPA) at edge surveillance units to generate compressed video embeddings, reducing transmission overhead while enabling early situational awareness. Via V2X links, these embeddings are transmitted to vehicle units equipped with a lightweight, small language model enhanced by a visual adapter, enabling it to perform visual-textual processing and respond to real-time traffic queries related to future collision predictions and environment states such as vehicle density. Results show that the model achieves an 89% F1 score for predicting the future collision events.