For autonomous vehicles driving on roads, future trajectories of surrounding traffic agents (e.g., vehicles, bicycles, pedestrians) are essential information. The prediction of future trajectories is challenging as the motion of traffic agents is constantly affected by spatial-temporal interactions from agents and road infrastructure. To take those interactions into account, this study proposes a Graph Attention Transformer (Gatformer) in which a traffic scene is represented by a sparse graph. To maintain the spatial and temporal information of traffic agents in a traffic scene, Convolutional Neural Networks (CNNs) are utilized to extract spatial features and a position encoder is proposed to encode the spatial features and the corresponding temporal features. Based on the encoded features, a Graph Attention Network (GAT) block is employed to model the agent-agent and agent-infrastructure interactions with the help of attention mechanisms. Finally, a Transformer network is introduced to predict trajectories for multiple agents simultaneously. Experiments are conducted over the Lyft dataset and state-of-the-art methods are introduced for comparison. The results show that the proposed Gatformer could make more accurate predictions while requiring less inference time than its counterparts.

Despite the advancements in the technologies of autonomous driving, it is still challenging to study the safety of a self-driving vehicle. Trajectory prediction is one core function of an autonomous vehicle. This study proposes an Attention-based Interaction-aware Trajectory Prediction (AI-TP) for traffic agents around the autonomous vehicle. With an encoder-decoder architecture, the AI-TP model uses Graph Attention Networks (GAT) to describe the interactions of traffic agents and Convolutional Gated Recurrent Units (ConvGRU) to carry out predictions. Based on the attention mechanism, the AI-TP model con-structs graphs from various traffic scenes to predict trajectories of different types of traffic agents. Traffic data from both the high-way (i.e., NGSIM) and urban road areas (i.e., ApolloScape and Argoverse) are used to evaluate the performance of the AI-TP model. Numerical results demonstrate that the AI-TP model requires less inference time and achieves better prediction accuracy than state-of-the-art methods. Specifically, the AI-TP model improves the performance by 36% on average with much less inference time on the NGSIM dataset, which shows the promise of predicting trajectories under various scenarios. The code of the AI-TP model will be available at https://github.com/KP-Zhang/AI-TP.

Accurate short-term passenger demand prediction contributes to the coordination of traffic supply and demand. This paper proposes an end-to-end multi-task learning temporal convolutional neural network (MTL-TCNN) to predict the short-term passenger demand in a multi-zone level. Along with a feature selector named spatiotemporal dynamic time warping (ST-DTW) algorithm, this proposed MTL-TCNN is quite qualified for the multi-task prediction problem with the consideration of spatiotemporal correlations. Then, based on the car-calling demand data from Didi Chuxing, Chengdu, China, and taxi demand data from the New York City, the numerical results show that the MTL-TCNN outperforms both classic methods (i.e., historical average (HA), v-support vector machine (v-SVM), and XGBoost) and the state-of-the-art deep learning approaches [e.g., long short-term memory (LSTM) and convolutional LSTM (ConvLSTM)] in both the single task learning (STL) and multi-task learning (MTL) scenarios. In summary, the proposed MTL-TCNN with the ST-DTW algorithm is a promising method for short-term passenger demand prediction in a multi-zone level.