Feasibility of clustering road user trajectories in complex scenes for automatic identification of common traffic activities

Proactive road safety analysis allows for the pre-emptive diagnosis of road safety issues without direct observation of traffic accidents by observing accident precursor events instead (i.e. "traffic conflicts"). This approach to road safety diagnosis is made possible with the collection and analysis of large quantities of high-resolution road user trajectory data acquired from video data automatically. However, several practical challenges with implementing this automation remain, including the automatic recognition of activity types in congested and complex scenes, particularly if the trajectory data is noisy. This activity recognition provides contextual information when observing traffic conflicts necessary for understanding specific causes of road safety issues and provides a better understanding of potential collision mechanisms.
Although this task can be performed manually, automation is sought for large-scale application of this technology as the manual task of performing activity recognition becomes cost-prohibitive. This project aims to achieve automated traffic activity recognition with a combination of trajectory clustering techniques and lane usage-learning heuristics from previously available road user classification (itself obtained from image recognition). Feasibility of this approach will be studied, including a sensitivity analysis of trajectory clustering in highly complex urban environments (e.g. intersections) and with lane type identification (road, sidewalk, bike path, etc.).

Paul St-Aubin
Faculty Supervisor: 
Liping Fu
Partner University: