Action Sequence Predictions of Vehicles in Urban Environments using Map and Social Context

This work studies the problem of predicting the sequence of future actions\nfor surround vehicles in real-world driving scenarios. To this aim, we make\nthree main contributions. The first contribution is an automatic method to\nconvert the trajectories recorded in real-world driving scenarios to action\nsequences with the help of HD maps. The method enables automatic dataset\ncreation for this task from large-scale driving data. Our second contribution\nlies in applying the method to the well-known traffic agent tracking and\nprediction dataset Argoverse, resulting in 228,000 action sequences.\nAdditionally, 2,245 action sequences were manually annotated for testing. The\nthird contribution is to propose a novel action sequence prediction method by\nintegrating past positions and velocities of the traffic agents, map\ninformation and social context into a single end-to-end trainable neural\nnetwork. Our experiments prove the merit of the data creation method and the\nvalue of the created dataset - prediction performance improves consistently\nwith the size of the dataset and shows that our action prediction method\noutperforms comparing models.\n