Abstract
As autonomous driving technology progresses, the need for precise trajectory prediction models becomes paramount. This paper introduces an innovative model that infuses cognitive insights into trajectory prediction, focusing on perceived safety and dynamic decision-making. Distinct from traditional approaches, our model excels in analyzing interactions and behavior patterns in mixed autonomy traffic scenarios. We introduce the Macao Connected Autonomous Driving (MoCAD) dataset as part of our contributions, which adds value to its complex urban driving scenarios. Our model represents a significant leap forward, achieving marked performance improvements on several key datasets. Specifically, it surpasses existing benchmarks with gains of 16.2% on the Next Generation Simulation (NGSIM), 27.4% on the Highway Drone (HighD), and 19.8% on the MoCAD dataset. Our proposed model shows exceptional proficiency in handling corner cases, essential for real-world applications. Moreover, its robustness is evident in scenarios with missing or limited data, outperforming most of the state-of-the-art baselines. This adaptability and resilience position our model as a viable tool for real-world autonomous driving systems, heralding a new standard in vehicle trajectory prediction for enhanced safety and efficiency.
Introduction: The Challenge of Mixed Autonomy
Predicting vehicle trajectories in real-world environments is complicated by "mixed autonomy," where self-driving cars (AVs) must share the road with unpredictable human drivers (HVs). Traditional deep learning models often rely solely on historical coordinates, lacking a "theory of mind" to understand human psychology. For instance, an aggressive driver might perceive a tight gap as safe and accelerate, while a conservative driver would view the same situation as dangerous and slow down.
This "cognitive gap" makes current AVs struggle with high-risk "corner cases" where human behavior dictates outcomes more than simple physics. Our research bridges this gap by teaching AVs to "feel" the road through cognitive-driven insights.
Methodology: Cognitive-Driven Architecture
Our model incorporates three novel modules designed to emulate human decision-making processes:
- Perceived Safety-Aware Module: This module moves beyond objective distance metrics to calculate a Quantitative Safety Assessment (QSA). By utilizing metrics like Time-To-Collision (TTC) and a novel Risk Tendency Index (RTI), the model quantifies how "dangerous" a scene feels to different driver types.
- Driver Behavior Profiling (DBP): Using Dynamic Geometric Graphs (DGG) and centrality measures, this module identifies driver styles (e.g., aggressive vs. conservative) in real-time without requiring manual labels.
- Priority-Aware Module: This module employs a spatial pooling mechanism to prioritize surrounding agents that have the most significant impact on the target vehicle's path.
These cognitive features are integrated into a Leanformer—an efficient Transformer variant—and processed by a Multimodal Decoder to generate probabilistic future trajectories.
Experimental Results and Robustness
The model was rigorously tested on the NGSIM, HighD, and our new MoCAD dataset. Key findings include:
- Superior Accuracy: The model achieved significant accuracy gains over state-of-the-art (SOTA) benchmarks, including a 27.4% improvement on the HighD dataset.
- Data Efficiency: Even when trained on only 25% of the available data, our model outperformed many existing models trained on full datasets.
- Resilience to Data Loss: In tests where 3 to 8 frames of history were intentionally removed to simulate sensor failure, the model maintained high performance. This suggests that the cognitive understanding of intent and risk allows the system to "fill in the blanks" when raw data is missing.
Conclusion
By integrating perceived safety and behavior profiling, our model establishes a new benchmark for trajectory prediction. Its ability to handle complex social logic and maintain robustness during data dropouts makes it a highly viable solution for the next generation of autonomous driving systems. Future work will explore integrating Large Language Models (LLMs) to further enhance the reasoning capabilities of AVs in complex topological environments.
Summary updated