Abstract
Intention prediction is a crucial task for Autonomous Driving (AD). Due to the variety of size and layout of intersections, it is challenging to predict intention of human driver at different intersections, especially unseen and irregular intersections. In this paper, we formulate the prediction of intention at intersections as an open-set prediction problem that requires context specific matching of the target vehicle state and the diverse intersection configurations that are in principle unbounded. We capture map-centric features that correspond to intersection structures under a spatial-temporal graph representation, and use two MAAMs (mutually auxiliary attention module) that cover respectively lane-level and exit-level intentions to predict a target that best matches intersection elements in map-centric feature space. Under our model, attention scores estimate the probability distribution of the open-set intentions that are contextually defined by the structure of the current intersection. The proposed model is trained and evaluated on simulated dataset. Furthermore, the model, trained on simulated dataset and without any fine tuning, is directly validated on in-house real-world dataset collected at 98 real-world intersections and exhibits satisfactory performance, demonstrating the practical viability of our approach.
Introduction: The Challenge of Intersection Complexity
Predicting a human driver's intent at an intersection is fundamental to safe autonomous driving. Traditional approaches often treat this as a "closed-set" classification problem, assigning labels like "left," "right," or "straight." However, real-world intersections rarely follow a standard four-way grid. They feature varying numbers of exits, irregular angles, and complex lane topologies that make simple labels ambiguous. For instance, an intersection with multiple left-turn lanes leading to different exits cannot be accurately modeled by a single "turn left" class.
This paper shifts the paradigm from classification to an open-set matching problem. By treating the map as a structured prior, the model matches the vehicle's state to specific map elements, allowing it to generalize to any intersection geometry, regardless of whether it was seen during training.
Methodology
The proposed framework leverages a map-centric approach that decomposes intention into two hierarchical levels: Goal (Exit) level and Lane level.
Spatiotemporal Graph Representation
The intersection environment is modeled as a graph where nodes represent exits, virtual lanes, and the target vehicle. This representation ensures the model is invariant to the number of lanes or the rotation of the intersection. Temporal dynamics are captured using Gated Recurrent Units (GRUs) to process the vehicle's movement history, effectively modeling the evolution of driver intent.
Map-Centric Feature Encoding
To avoid the computational overhead of high-dimensional rasterized images, the model utilizes curvilinear coordinates relative to lane centerlines and exit points. These features are spatially invariant, focusing on the vehicle's position relative to the road structure rather than absolute global coordinates.
Mutually Auxiliary Attention Modules (MAAMs)
The core of the architecture consists of two attention modules designed to exploit the physical constraints between lane choice and exit choice.
- Lane Attention Module: Refines lane-level probabilities by incorporating goal-level information.
- Goal Attention Module: Aggregates lane-level features to improve the prediction of the final exit.
This bi-level constraint mechanism ensures that the predicted intentions are physically plausible (e.g., a vehicle in a dedicated left-turn lane is unlikely to take a right exit).
Experimental Evaluation
The model was trained and tested using both simulated and real-world data:
- Simulated Dataset: 89,000 sequences from 145 distinct intersection layouts in Shanghai, providing ground-truth for both lane and goal intentions.
- Real-World Dataset: 6,346 trajectories from 98 intersections. Notably, 47 of these intersections were entirely new to the model, providing a rigorous test for zero-shot generalization.
Key Results and Insights
- Generalization: The model achieved a 98.8% recall for goal intention in simulation. More impressively, it maintained a 95.1% recall on unseen real-world intersections without any fine-tuning, demonstrating robust zero-shot performance.
- Efficiency: The graph-based approach is highly efficient, with an average inference time of 0.3ms per target. This is significantly faster than CNN-based methods, making it ideal for real-time deployment on automotive hardware.
- Robustness: Qualitative analysis indicates the model remains accurate even with noisy map data or missing lane centerlines.
Conclusion
By framing intention prediction as an open-set matching problem and utilizing mutually auxiliary attention, this work provides a scalable and efficient solution for complex intersection navigation. Future work will explore extending this framework to highway scenarios and integrating it into multi-modal trajectory prediction systems.
Summary updated