Semantic Reasoning over Scene Graphs for Probabilistic Prediction of Traffic Agents using Graph Neural Networks

This thesis primarily focusses on developing a generic graph representation of traffic scenarios that will be used to classify attributes of traffic agents using Graph Neural Networks.

Completed Master Thesis

Autonomous driving requires accurate prediction of future states of other traffic agents in order to navigate safely. Deep learning techniques can be used as data-driven solution for the prediction problem, because the required labels can be derived from future states of traffic agents. Besides directly predicting agent trajectories, one can also estimate the underlying semantic intent of each agent. A causal relationship exists from semantic intent to resulting trajectory. Because no complete dataset for agent intents is available, this work will focus on classifying agent attributes in general. We aim to develop a scene graph that represents the real-world setting and contains information about static and dynamic entities and their relations in the scenario. As a novelty, we will research towards a holistic representation that contains a larger set of input features, including information from the HD map. A Graph Neural Network (GNN) will be trained to process the proposed scene graph in order to predict attributes of traffic agents. An iterative approach is chosen where features are systematically integrated into the scene graph structure. In the evaluation phase, the trained GNN models are then compared to existing baseline approaches. As part of future work, the approach can be extended to predict trajectory endpoints. Optionally, an ablation study could quantify the effect of individual inputs on the overall performance.



To the top of the page