1667162 - Real-Time Traffic Perception on IoT: Conflicts Prediction and Warning Systems at Intersections

Abstract: Intersection safety, with its intertwined traffic dynamics, emerges as a global concern, especially in urban environments where collision rates peak. Although adept at post-collision data processing, prevailing sensing systems falter in preempting potential conflicts, thus limiting the enhancement of intersection safety. The deficit in real-time collision data further accentuates this issue. Addressing this critical gap, our research introduces a Predictive Alert and Sensing System (PASS) designed to identify potential traffic conflicts up to three seconds in advance and proactively alert road users, reducing the risk of collisions. 

The PASS is structured around three core modules. Firstly, the Video-based Perception Module (VPM) is equipped with powerful Edge-AI technologies to detect and track different kinds of road users at the intersection to extract their trajectories. Secondly, the Trajectory Prediction Module (TPM) enables physically informed trajectory forecasting, assimilating crucial parameters such as speed limits, lane positioning, signal timings, and prevailing weather conditions to ensure high prediction fidelity, especially given the erratic movement dynamics at intersections. Finally, the Collision Prediction Module (CPM) pinpoints imminent collisions by comparing the anticipated arrival times of intersecting trajectories. Significantly, the entire system is optimized for edge computing, epitomized in its deployment on Jetson Nano, equipped with a robust quad-core ARM A57 64-bit CPU and a 128-core Maxwell GPU. Field tests at two bustling intersections in Bellevue, WA, underscored its prowess. Predicted trajectories over an impending three-second window achieved accuracies of 95%, 92%, and 84% for each successive second. The potential conflict prediction accuracy was benchmarked at approximately 92%, adeptly identifying both actual and near-miss collision scenarios. Furthermore, its stable 30 FPS processing speed shows the system's operational efficiency. Coupling internet and local communication avenues, it swiftly (in less than 200ms) transmits warning notifications to proximate road users and relevant transport agencies, significantly enhancing the proactive mitigation of potential accidents.

The contributions of the research can be summarized into four aspects: 1) Architecture: We introduce a cutting-edge predictive sensing framework, Predictive Alert and Sensing System (PASS), designed to foresee and mitigate potential intersection collisions, thereby enhancing traffic safety. 2) Methodology: The study brings to the fore three pioneering, physically-informed Edge AI techniques, ensuring precise and efficient traffic perception and prognostication. 3) Edge-Adaptation: The proposed PASS is meticulously fine-tuned for edge computing. This bolsters its efficiency and fortifies cybersecurity measures, safeguarding road users' privacy. 4) Practical Implementation: Our system has been piloted at two Bellevue intersections as a testament to its potential. Preliminary results underscore the promise and efficacy of the PASS system.

In conclusion, this research proposed the PASS system regarding traffic safety, intricately engineered using pioneering Edge AI techniques to proactively predict and mitigate potential traffic conflicts, bolstered by its practical efficacy and robust cybersecurity measures, manifesting a significant stride towards enhanced urban traffic safety.