Recursive prediction network: surrounding vehicle trajectory prediction with future trajectory feedback

Abstract

The perception and control technology for autonomous vehicles shows significant advances owing to the deep learning techniques. On the other hand, the prediction technology for the future motion of surrounding vehicles remains a challenging problem due to the complexity of vehicle motion. Accidents of autonomous vehicles from Google and Uber can also be attributed to inaccurate predictions. The motion of a vehicle is not only determined by the intention of the driver but is influenced by various interactions with other vehicles. In order to tackle this problem, various approaches for vehicle motion prediction have been developed, but these approaches failed to predict future motion accurately because they use only current and historical information. These future interactions are difficult to predict from current and historical information. Therefore, in this thesis, a vehicle trajectory prediction network that can consider not only current and past interaction but also future interactions, which have not yet happened, is proposed. To this end, a recursive structure, which uses the output of the network as input again, is added to a Long Short Term Memory (LSTM) based encoder-decoder model. With this recursive structure, the model can use the predicted future trajectory together with the current and past information of surrounding vehicles, and this makes the model possible to predict the future trajectory considering future interaction. The proposed method can be expected to play a vital role in moving forward to fully autonomous driving by improving the reliability of prediction technology.