Decision-Making Strategy of Automated Vehicles in Heterogeneous Scenarios Based on Attention Networks

Autonomous driving decision-making ability is a key technology for intelligent vehicles. Traditional rule-based autonomous driving decision-making methods set corresponding rules in relatively simple single scenarios, which are not suitable for complex and diverse multi-source heterogeneous driving scenarios. The Learning methods based on Spatio-Temporal Graph Neural Networks (ST-GNNs), unlike recurrent networks or convolution networks, can extract the complexity of spatial relationships and temporal dependencies and output driving actions for a period of time in the future. However, most ST-GNNs used for autonomous driving decision-making only consider vehicle nodes, ignoring the heterogeneous node attributes that record rich scenario semantic information, which have not been fully utilized to guide better graph structure learning. But these semantic information are equally crucial for autonomous driving decision-making. To couple with these problems, we propose an Spatio-Temporal Heterogeneous graph Attention Network, namely ST-HANet, to first construct a dynamic heterogeneous graphs to represent a multi-source heterogeneous driving scenarios. Furthermore, for spatial dimension, via mining heterogeneous graph information through heterogeneous graph attention network, and fuse the mined relationship representations with ego vehicle node information through cross attention mechanisms. Next, for temporal dimension, capturing the temporal dependencies of the sequences through the temporal attention mechanism. Moreover, we conducted open-loop and closed-loop testing of the model based on the CARLA simulator, and compared it with state-of-the-art methods. The experimental results demonstrate that this method has higher accuracy and better decision-making performance than existing methods.