EGL: Efficient Graph Learning with Safety Constrains for Heterogeneous Trajectory Prediction

Weijie Lian, Yuze Wang, Ximu Zeng, Shuncheng Liu, Yuyang Xia, Huiyong Tang, Han Su

Abstract: Autonomous vehicles follow predicted trajectories to avoid obstacles and to drive safely. Any trajectory prediction algorithms that ignore the importance of safety constraints may lead to collisions in the real world, especially in a complex environment. In this paper, we tackle this problem with a novel trajectory prediction algorithm called Efficient Graph Learning (EGL), which utilizes road network information, self-attention mechanism, and safety constraints. In EGL, in order to obtain the road network information, real-world images are embedded in Convolutional Neural Networks (CNN), and then a Global Graph is constructed to present road network information and vehicle characteristics. The Global Graph is sent to an encoder-decoder model, which contains a Graph Attention Network(GAT) layer and Long Short-Term Memory (LSTM) layer in both the encoder and decoder. When the LSTM in the decoder outputs the results, EGL can generate all the predicted trajectories simultaneously. We evaluate EGL on some datasets. Experiment results show that EGL performs well in heterogeneous datasets, and outperforms other state-of-the-art methods with improving average displacement error (ADE) and final displacement error (FDE) by 7\(\%\) and 27\(\%\) respectively.