Research Overview

Comprehensive Traffic Environment Perception with Edge Intelligence

Motivation: The physical traffic environment plays a crucial role in influencing the traffic system. Unfortunately, neglecting these essential environmental aspects can lead to severe repercussions, especially under adverse weather conditions, including a decline in accuracy and the emergence of anomalies. 

Research Summary: To mitigate the uncertainties arising from the traffic environment, we proposed an Edge-based Multi-task Safety-oriented Environmental (Edge-MuSE) sensing system. Edge-MuSE is designed to provide an extensive perception of the traffic environment, encompassing temperature, humidity, image dehazing, visibility measurement, and road surface condition classification. This innovative system is tailored for comprehensive, efficient, and cost-effective real-time monitoring, enhancing the reliability and safety of traffic management systems.

Bias Mitigation in Uneven Distributions by Integrating Physical Information

Motivation: Given the inherently uneven and long-tailed distribution of road users in transportation scenarios, deep learning technologies can unintentionally amplify biases, resulting in diminished accuracy for less-represented classes like children and wheelchair users. Accurately quantifying and mitigating the bias of deep-learning-empowered perception systems to address this equity issue is therefore a significant research focus for daily traffic services and applications.

Research Summary: In response, we developed a trustworthy perception system based on few-shot and representation learning algorithms, DEbias Traffic Object Recognition (DETOR), for road user and traffic facility detection and classification. By employing DETOR, the residual neural network achieved a 290% boost in accuracy for minority transportation object detection. The research developed a pioneering trustworthy, transportation-adaptive, de-biased traffic perception system, paving the way for more promising future transportation services and applications.

Privacy-Preserving Cooperation Systems for Cross-Camera Network Monitoring

Research Summary: To track objects across multiple cameras and help public agencies collect link travel time and speed information without any sensitive privacy information, an Edge-empowered Cooperative Multi-camera Sensing (ECoMS) System is proposed. ECoMS presents a novel algorithmic and edge-server cooperative framework that pushes edge computing and multi-camera re-identification workflows to serve traffic sensing on a cooperative IoT architecture. On the algorithm side, ECoMS proposes a lightweight edge-based computer vision framework for real-time vehicle detection, tracking, and representation selection. By sending only the objects’ representations to the edge server, high-bandwidth data transmission and heavy post-processing can be eliminated. Further, a hierarchical clip-based deep vehicle re-identification framework is integrated into ECoMS and significantly outperforms other state-of-the-art methods with 92.43% Rank-1 accuracy. Finally, a cooperative cross-camera traffic information estimation framework based on kernel density estimation with a kernel smoother is implemented, producing precise and reliable cross-region traffic parameters (link travel time, speed, and area OD) and distributions (less than 1.01 KL distance).

Synchronized Infrastructure System with Cooperative Engagement for Dilemma Zone Mitigation

Research Summary: The dilemma zone is a significant safety concern in transportation, where existing infrastructure systems often lack the real-time processing and communication capabilities to mitigate the heightened risk of crashes caused by driver hesitation at signal changes. This research introduces an edge-server-based platform that unifies surveillance cameras, traffic signal controllers, and road users through a cooperative framework of perception, communication, and control to mitigate dilemma zone risks. The platform is structured into three core modules: a sensing module for instantaneous vehicle detection and tracking; a data processing module for trajectory data analysis and forecasting; and a communication module that facilitates immediate interaction with infrastructure and motorists. Deployed at two pioneer intersections in downtown Bellevue for one week, the system detected 1,076 dilemma events with 92.84% accuracy. It is the first connected infrastructure system to integrate sensing, data processing, and information dissemination for self-operating signal services.