| Abstract |
High-quality trajectory prediction is crucial for autonomous vehicles with various applications like driving safely and planning traffic trajectory in smart cities. In order to achieve high-precision trajectory prediction, not only the historical trajectory information of the target needs to be considered, but the influence of vehicle interaction on the future trajectory is also extremely important. However, (1) with the increase of the input historical trajectory sequence, it is cause the important information in the feature sequence to be overwritten or lost; (2) It is also a challenge to properly model the interaction between vehicles. In this paper, we propose a Self-Attention Convolutional Social pooling LSTM (SACS-LSTM) vehicle trajectory prediction model. It uses the self attention mechanism to make important information in the historical sequence not easily lost, and uses the self attention mechanism to improve the modeling accuracy of the convolutional social pooling layer for vehicle interaction information, so that the model can achieve the purpose of accurate modeling. We use the NGSIM data set to evaluate the model, and the results vertify that the predicted accuracy within 5s outperforms the baseline models. © 2021 ACM. |
