Classifying Road Users in Urban Scenes Using Movement Patterns

This paper describes an automated classification approach to road users. The main motivation behind road-user classification in the context of safety stems from the necessity to learn traffic scenarios and understand patterns within each road-user class. The end goal in the analysis is to identify and learn scenarios that may contribute to hazards in traffic conditions. The classification relies on video data (movement trajectories) collected in urban intersections. The approach is based on the discrimination of the shapes of the speed profiles of each road-user type, more precisely, the discrimination between the speed movement patterns of vehicles and the ambulatory characteristics of pedestrians. The collected movement-trajectory data are represented as time series. The classification is performed using singular value decomposition and reconstruction of the time series. Two complementary methods are proposed based on the quality evaluation (correlation score) of the reconstructed trajectories. In the first method, a threshold-based decision procedure is applied. This approach is complemented in the second method by a semisupervised classification procedure guided by movement prototypes. The approach is validated on real-world data collected in Oakland, California. A correct classification of around 90% was achieved using both methods.