Autonomous driving is a dynamic field of research that has garnered significant attention and effort from numerous researchers. Its progress has been greatly propelled by the remarkable success of deep learning methods. Nevertheless, there remain enduring obstacles to overcome in order to effectively implement autonomous driving in intricate real-world scenarios. This thesis aims to tackle two prominent challenges in this domain: trajectory prediction and uncertainty quantification. The first contribution of this thesis is to introduce a novel framework for trajectory prediction that specifically addresses scenarios where driver behavior is heavily influenced by interactions. The framework leverages recurrent neural networks and a non-local block to learn trajectory representations. Additionally, a graph neural network is proposed to effectively model and capture the intricate interactions between vehicles. This contribution enhances our understanding and prediction capabilities in dynamic driving environments where interactions play a crucial role. The second contribution of this work focuses on providing an alternative approach to using expensive HD maps by incorporating historical local behaviors. By harnessing the valuable insights derived from historical local behaviors, which encompass lane-specific information, the framework effectively integrates this data into the prediction pipeline using an attention mechanism. This approach offers a cost-effective solution that leverages the power of historical data to enhance the accuracy and reliability of trajectory prediction, thereby reducing the reliance on resource-intensive HD maps. The third significant contribution is the development of a novel framework for uncertainty quantification. This framework effectively captures and models uncertainty by mapping specific attributes of prediction models, without the need for modifying the models’ structure or retraining them. This contribution enhances the applicability and robustness of autonomous driving systems by enabling a more comprehensive understanding of the uncertainties associated with prediction models. The final contribution of this research builds upon the third contribution by introducing a Gaussian Mixture Model (GMM) to accurately describe the distribution of specific attributes within prediction models for modeling uncertainty. This extension enhances the framework’s ability to capture and represent uncertainty in a more nuanced manner. Additionally, the method is combined with the Gaussian Sum Filter, establishing a more robust estimation pipeline that effectively combines estimation techniques with prediction models.