Pedestrian crossing prediction using multiple context-based models

In inner-city, most vehicle-pedestrian collisions occur when a pedestrian is crossing the road and the driver does not see or pay attention to him. Current ADAS (advanced driver assistance systems) warn the driver or apply the brakes shortly before the collision, but in some situations the collision cannot be fully avoided because most systems react only when the pedestrian is already in front of the vehicle. To fully avoid a collision, a driver should be warned earlier. Behavior prediction is a solution that can be used to warn a driver before the pedestrian starts crossing. In this paper, we propose a generic context based model to predict crossing behaviors of pedestrians in inner-city. We will show that our model provides accurate prediction at an early time. However, there are specific locations such as zebra crossings, where based on expert driving experience, one would expect that a prediction can be done even earlier. Therefore, we have developed an additional specific model fitted to the context of zebra crossings. The experiments show that this model produces both, better and earlier predictions in this specific context. Because our goal is to build a generic crossing prediction system, we finally apply the framework of the `Context Model Tree' to combine the two models. We demonstrate that this multi-model system is well suited to provide early predictions for realistic data, including both, generic inner-city situations and zebra crossings.