Towards Safer and Understandable Driver Intention Prediction

TL;DR

This work addresses interpretability in driver intention prediction (DIP) by introducing DAAD-X, a multimodal explainable driving action dataset with in-cabin eye-gaze and out-cabin ego-vehicle explanations, and by proposing the Video Concept Bottleneck Model (VCBM). VCBM combines dual video encoders with Learnable Token Merging (LTM) and Localised Context Bottleneck (LCBM) to produce spatio-temporally coherent, human-understandable explanations while predicting driving maneuvers, framed by a joint loss over maneuver and explanations. Experiments show transformer-based backbones coupled with LTM and LCBM outperform baselines in explanation fidelity and action accuracy, and reveal the critical role of gaze cues and temporal context for interpretable predictions. The work also introduces a multi-label t-SNE visualization to illustrate the disentanglement and causal relationships among explanations, contributing to safer, more transparent autonomous driving systems.

Abstract

Autonomous driving (AD) systems are becoming increasingly capable of handling complex tasks, mainly due to recent advances in deep learning and AI. As interactions between autonomous systems and humans increase, the interpretability of decision-making processes in driving systems becomes increasingly crucial for ensuring safe driving operations. Successful human-machine interaction requires understanding the underlying representations of the environment and the driving task, which remains a significant challenge in deep learning-based systems. To address this, we introduce the task of interpretability in maneuver prediction before they occur for driver safety, i.e., driver intent prediction (DIP), which plays a critical role in AD systems. To foster research in interpretable DIP, we curate the eXplainable Driving Action Anticipation Dataset (DAAD-X), a new multimodal, ego-centric video dataset to provide hierarchical, high-level textual explanations as causal reasoning for the driver's decisions. These explanations are derived from both the driver's eye-gaze and the ego-vehicle's perspective. Next, we propose Video Concept Bottleneck Model (VCBM), a framework that generates spatio-temporally coherent explanations inherently, without relying on post-hoc techniques. Finally, through extensive evaluations of the proposed VCBM on the DAAD-X dataset, we demonstrate that transformer-based models exhibit greater interpretability than conventional CNN-based models. Additionally, we introduce a multilabel t-SNE visualization technique to illustrate the disentanglement and causal correlation among multiple explanations. Our data, code and models are available at: https://mukil07.github.io/VCBM.github.io/

Figures (8)

• Figure 1: Illustration of an AD scenario for the DIP task. An AD system may intend to take a left turn while encountering a parked or slow-moving vehicle at the turn. Existing DIP models, lacking HCI understanding, might fail to anticipate the obstacle, leading to a potential collision. The proposed interpretable model, VCBM, enhances safety by enabling the ego-vehicle to explain its intended actions, anticipate obstacles more effectively, and adjust maneuvers accordingly. This results in safer and more transparent decision-making.
• Figure 2: Driving video annotation statistics of DAAD-X dataset. Illustrating the distribution of (left) ego-vehicle explanations and (right) eye-gaze explanations across different maneuver actions. Details of the full explanation annotation are provided in the Supplementary Material. Zoom in for better clarity.
• Figure 3: Overall architecture of the proposed VCBM. The dual video encoder first generates the spatio-temporal features (tubelet embeddings) for the ego-vehicle and gaze input sequence video pair. These tubelets are concatenated along the channel dimension and fed into the proposed learnable token merging block to produce $K$-cluster centers based on composite distances. These clusters are then fed into a localised concept bottleneck to disentangle and predict the maneuver label and one or more explanations to justify the maneuver decision.
• Figure 4: VCBM merges relevant features across frames $(z_{c_j})$ and assigns explanations. Blue represents merged traffic features, orange denotes merged traffic light features, and arrow thickness indicates prediction confidence.
• Figure 5: Variants of gaze input. The driver's view video is processed in the following way to show the best way to represent gaze without affecting spatial features. The gaze cropped variant ($R=350$) produces the best quantitative results.