A Concept-Based Explainable AI Approach to Action Recognition in Autonomous Driving

In recent years, the field of autonomous driving has garnered increasing interest. Beyond ensuring the reliable operation of autonomous vehicles (AV), increasing attention has been directed towards understanding how these vehicles make decisions and predictions. Enhancing user trust in AVs requires developing situational awareness and decision-making processes that approximate human-level performance. Incorporating these capabilities in AV not only improves safety but also facilitates more intuitive and reliable human-machine interactions. Achieving this requires AI models capable of interpreting and responding to complex road scenarios in a manner similar to an experienced human driver. The ability to accurately predict and respond to dynamic traffic conditions is critical for the success and widespread adoption of AVs. Given the safety-critical nature of autonomous driving, interpretable predictive models are essential for building user trust. This thesis focuses on integrating methodologies that enhance the explainability of AV decision-making, making the reasoning behind actions more understandable and transparent. The initial part of this work reviews the state of the art by analyzing various frameworks designed to collect, integrate, and fuse heterogeneous data from multiple sensors — such as cameras, GPS, and LiDAR— to support optimal decision-making and control commands. This project focuses exclusively on data from a centrally positioned, front-facing camera aligned with the vehicle’s direction of travel. Accordingly, the ROAD dataset was selected as the most suitable for this case study. The dataset is particularly well-suited for analysing road scenarios and developing models that improve situational awareness and decision-making in autonomous vehicles, stems form its annotation-based structure, which utilizes Road Events to describe key elements within a scene. Several neural network architectures are evaluated to compare the generalization capabilities of black-box models with those of explainable approaches. The main goal is to introduce a level of interpretability into the decision-making process, in order to better understand the rationale behind the actions taken by the ego-vehicle. This is achieved by applying a Concept Bottleneck Model (CBM) approach, which enables intermediate predictions of human-interpretable concepts. To identify high-level concepts within the images the 3D-RetinaNet model was employed. This model enables the extraction of concepts, i.e., key elements from the scene, similar to the way a human driver would reason when deciding on an action. Based on these extracted concepts, various neural networks were subsequently tested to predict the autonomous vehicle’s decision-making behaviour, leading to a model with enhanced interpretability by grounding decisions in identifiable scene elements. While explainability provides valuable insights into the decision-making process, it can sometimes lead to a reduction in predictive performance. Therefore, it is important to find an appropriate trade-off between model interpretability and performance.