Lane detection by trajectory clustering in urban environments

Extraction of road geometry and vehicle motion behaviour are important for the semantic interpretation of traffic flow patterns, as a component of an intelligent vision-based traffic surveillance system. This paper presents a method for computing the location of traffic lanes by clustering vehicle trajectories. It employs a novel trajectory detection and clustering algorithm based on a new trajectory similarity distance. Moving vehicles are detected against a background estimated using a self-adaptive Gaussian mixture model (SAGMM), and fitted by a simple wireframe model. The vehicle is tracked by a Kalman filter using a landmark feature that is close to the road surface. The centre line of each traffic lane is computed by clustering many trajectories. Estimation bias due to vehicle lane changes is removed using Random Sample Consensus (RANSAC). Finally, atypical events associated with vehicles departing from the normal lane behaviours (e.g. lane changes) are detected.