H. Cheng (Hao)

Currently, I am a recipient of Marie Skłodowska-Curie Actions (MSCA) European Postdoctoral Fellowship with the 101062870 - VeVuSafety project that aims to learn the interactions between vehicles and vulnerable road users to facilitate safer Intelligent Transportation Systems and Autonomous Driving.

My research interest includes deep learning of road user behavior modeling in Intelligent Transport Systems and Autonomous Driving and Safety Analysis between vehicles and vulnerable road users.

Publications

Traffic safety is the fundamental criterion for vehicular environments, and for the development of many Artificial Intelligent (AI) systems like self-driving cars and robot navigation systems. Unfortunately, statistics show that direct interactions between vehicles and vulnerable road users (VRUs, e.g., pedestrians and cyclists) still lead to serious injuries and fatalities. From 2010 to 2019 in the EU countries, pedestrians and cyclists account for about 20% and 9% respectively to the annual number of fatalities; Among different types of VRUs, pedestrians are the most vulnerable road users at intersections and require the most protection[2]. For instance, a vehicle driver may fail to recognize an approaching pedestrian at an intersection due to block of views, and some pedestrians may not pay attention to the upcoming traffic while talking to friends or looking at mobile phones. If these dangers are automatically detected in advance, a prompt signal can be sent by a traffic safety alarming system to prevent a potential accident. Moreover, the foreseeable advent of autonomous driving and mobile robot applications highly depends on AI systems that can correctly predict other road users' behaviour, i.e., their intended trajectories in the next seconds. Such systems are required to automatically process the input data extracted from the involved road users and the environment to generate desirable output that represents how they interact and move, to guarantee safe decisions to be made in the following steps.

However, there are some urban environments making detection and prediction of road users' behaviour particularly challenging, e.g., temporarily shared spaces of intersections for vehicle turning or shared spaces as a traffic design[1]. Both types of environments allow direct interactions between vehicles and VRUs with no time or space segregation for different transport modes. Under such weakened traffic rules and due to mutual influence, the road users’ behaviour tends to be multimodal and stochastic, such as stopping, accelerating, or turning. Often, a road user may have to change direction or travel speed to cope with a sudden change of the motion from others. In such situations, the human road user can subconsciously anticipate others' possible behaviour based on the information of road geometry, transport mode, distance, travel speed, orientation, and even other acoustic and facial information, to adjust her or his own movement accordingly. To the contrary, it is still a big challenge for AI systems to perceive the environment and automatically learn road users' behaviour in such kind of situations, and at the same time to protect road users’ personal information while processing massive traffic data.

Organisations

• Faculty of Geo-Information Science and Earth Observation (ITC), Department of Earth Observation Science (EOS)