For autonomous vehicles to drive safely, it is crucial to predict motion of other vehicles as well as to detect. The motion of the vehicle is affected by the location of other vehicles, road environments, and vehicle dynamics. In this paper, we propose a deep learning-based network that combines convolutional neural network (CNN) and Sequence-to-Sequence (Seq2Seq) for trajectory prediction. In order to encode the location of other vehicles and road environment information, raster images of surrounding vehicles and High-definition (HD) maps are taken as input of CNN. Also, the vehicle’s history positions and CNN features are taken as input of Seq2Seq since it implies vehicle’s motion and dynamics. Therefore, CNN extracts the traffic context and Seq2Seq encodes the vehicle’s motion and predicts the future position. We evaluate our method on Lyft dataset and show the prediction accuracy is improved compared to other methods. When the past 1 s path is given and the future 3 s path is predicted, the mean square error for the final horizon is 2.15, which is almost twice less than GRIP. In addition, our method runs at 25 ms.
Vehicle Trajectory Prediction with Convolutional Neural Network and Sequence-to-Sequence
Lect.Notes Mechanical Engineering
05.08.2021
6 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
SEQUENCE-TO-SEQUENCE PREDICTION OF VEHICLE TRAJECTORY VIA LSTM ENCODER-DECODER ARCHITECTURE
British Library Conference Proceedings | 2018
|Europäisches Patentamt | 2023
|