Vehicle Trajectory Prediction in Roundabout Based on the Joint Learning of Taillight State and Historical Trajectory

In complex urban road environments, especially at junctions such as roundabout and crossroads, predicting the future behaviour of surrounding vehicles can provide advanced awareness of possible hazards, thus helping decision-making and planning modules to act more efficiently and safely. Therefore, trajectory prediction is of great significance to the autonomous vehicles. However, trajectory prediction faces the problem of 'unclear intention', 'uncertain right-of-way' and 'frequent vehicle interaction' in interactive scenarios like roundabout. Therefore, this paper proposes a taillight state and historical trajectory joint learning method based on multi-head attention (MHA [15]) and LSTM [17] for vehicle trajectory prediction in roundabout. Specifically, the historical trajectory and taillight state sequence are encoded by MHA and LSTM respectively and concatenated together. On the condition of the current trajectory of ego vehicle, the feature information is input into a decoder LSTM network to predict the future trajectory of two target vehicles. Our method is trained and tested in the Carla [16] simulation environment and gets considerable results. Considering the error of taillight recognition in practice, the model is also tested under different taillight recognition accuracy and proven to be robust and practical. The dataset is published on: https://pan.baidu.com/s/1DWaWNbdUpj6UjYQNicqFDg?pwd=49jy.