An accurate Spatial Temporal Graph ATtention Network for pedestrian trajectory prediction

Pedestrian trajectory prediction has broad applications to target tracking and autonomous driving. Although current research has gradually shifted from handcrafted-based approaches to deep learning-based approaches, existing predictors sometimes produce wrong future trajectories, and their performance is still unsatisfied. This paper improves the Spatial Temporal Graph ATtention Network (STGAT) for trajectory prediction by adding two Multi-Layer Perceptron (MLP) modules and three Gated Recurrent Units (GRUs) to the original predictor. Specifically, the model uses an MLP to process the position and velocity of a pedestrian to obtain high-dimensional embedding features. Then, to improve prediction accuracy, the model uses a GRU and a Long Short-Term Memory module (LSTM), i.e. GRU-LSTM, to obtain the motion features of the individual pedestrian. Next, the interaction information between the target pedestrian and his/her neighbors is captured by a Graph ATtention (GAT) module. Here, to reduce the redundant information in the GAT, another MLP is used to map the high-dimensional features to low-dimensional ones. Next, the second GRU-LSTM further encodes the interaction features from the GAT. Finally, the third GRU-LSTM serves as a decoder to give the future trajectory of the target pedestrian. To generate multiple socially acceptable prediction results, the model adopts the variety loss. Test results on the public ETH and UCY datasets illustrate that the proposed model outperforms the state-of-the-art predictors such as Social-LSTM, Social-Attention, CIDNN, and etc. Moreover, the ablation tests show that by using the MLPs and GRUs, the ADE and FDE of the model are lower than the STGAT and Social-GAN, further proving the benefits of these new adding modules to the predictor.