Project description  

Traffic safety is the fundamental criterion for vehicular environments and many artificial intelligence-based systems like self-driving cars. There are places, e.g., intersections and shared spaces, in the urban environment with high risks where vehicles and vulnerable road users (VRUs) such as pedestrians and cyclists directly interact with each other. By advancing state-of-the-art artificial intelligence methodologies, this project VeVuSafety aims to build a privacy-aware deep learning framework to learn road users’ behavior in various mixed traffic situations for the safety of vehicles and VRUs. VeVuSafety proposes a 3D environment model based on a 3D point cloud for privacy protection — private information like license plates and faces is anonymized. Then, within this environment model, an end-to-end deep learning framework using camera data will be built for multimodal trajectory prediction, anomaly detection, and potential risk classification based on deep generative models such as the Variational Auto-Encoder. Additionally, an active privacy mechanism will also be adopted by application of the differential privacy mechanism to help the deep learning models prevent model-inversion attacks. Moreover, the framework’s generalizability will be investigated by exploring the Normalizing Flows approach for domain adaption. The framework’s performance will be validated at different intersections and shared spaces using real-world traffic data. Besides road user safety and privacy, VeVuSafety can help traffic engineers and city planners to better estimate the design of traffic facilities in order to achieve a road-user-friendly urban traffic environment.