Graph Neural Networks in ADAS: Architectures, Datasets and Common Approaches

Graph neural networks (GNNs) are neural networks that can be applied directly to graph datasets. This type of neural network can easily capture interactions between nodes in a graph and predict the connections between them. Recently, there has been an increase in the popularity of GNNs and their various applications and architectures, including graph convolution networks, graph attention networks, and graph recurrent neural networks. The enthusiasm surrounding GNNs stems from their capacity to operate effectively on unorganized data, which is common in real life. Autonomous vehicles (AV) and advanced driver assistance systems (ADAS) are a promising field of GNN applications. This field has a lot of interesting and challenging problems such as action recognition, trajectory prediction, and crossing attention. The objective of this study is to consolidate published works on GNN-based methods for estimating human behavior. Three viewpoints are considered: GNN architectures, benchmarking datasets, and common approaches. Then we end up by summarizing the feedback and observations we received during our research for future directions and research opportunities.