Abstract
Predicting the trajectories of human-driven vehicles (HVs) is a critical challenge for autonomous vehicles (AVs) operating in mixed autonomy environments. Traditional deep-learning approaches often treat trajectory prediction as a pure data-fitting problem, ignoring the underlying psychological factors that drive human behavior. This paper presents a novel cognitive-driven model that incorporates perceived safety and driver psychology. By modeling drivers' risk tolerance and social influence through a Dynamic Geometric Graph and a lightweight Leanformer architecture, the proposed system achieves significant performance gains—up to 27.4%—over state-of-the-art methods. Furthermore, the model demonstrates exceptional robustness to missing sensor data and high data efficiency.
Introduction: The Challenge of Mixed Autonomy
In mixed autonomy environments, AVs must navigate alongside human drivers whose actions are often influenced by subjective perceptions of safety and individual driving styles. Most current prediction models rely solely on historical coordinates, lacking a "theory of mind" to understand the motivations behind a sudden swerve or brake.
As shown in Figure 1, an aggressive driver might perceive a two-second time-to-collision as safe and choose to accelerate, while a conservative driver would find the same situation dangerous and slow down. Incorporating these cognitive insights is essential for developing AVs that can safely and naturally interact with human traffic.
Methodology: A Multi-Module Cognitive Architecture
The proposed model employs a multi-module architecture designed to emulate human decision-making processes:
- Perceived Safety-Aware Module: This core component combines quantitative safety metrics (like Time-To-Collision) with driver behavior profiling. It utilizes a Dynamic Geometric Graph (DGG) to treat traffic scenes as social networks, using centrality measures to identify a driver's influence and behavioral signature.
- Priority-Aware Module: This module recognizes the varying importance of surrounding vehicles based on their spatial positioning, encoding these relationships into high-dimensional vectors.
- Interaction-Aware Module (Leanformer): To ensure real-time efficiency, a lightweight "Leanformer" architecture fuses safety, behavior, and priority data.
- Multimodal Decoder: Utilizing a Gaussian Mixture Model (GMM), the decoder predicts multiple potential future trajectories with associated confidence scores, accounting for the inherent uncertainty in human behavior.
Experimental Results and Robustness
The model was evaluated on three major benchmarks: NGSIM (US highways), HighD (German highways), and MoCAD (urban Macao). The results demonstrate a significant leap in accuracy, with improvements of 16.2% on NGSIM, 27.4% on HighD, and 19.8% on MoCAD compared to existing state-of-the-art models. The improvements were most noticeable in long-term forecasts (4–5 seconds out), where understanding driver intent becomes more critical than just tracking momentum.
A key finding is the model's robustness to "data gaps." In tests where sensor frames were intentionally dropped to simulate real-world occlusions or failures, the cognitive-driven features allowed the model to maintain superior performance. Additionally, the model exhibited high data efficiency, outperforming baselines even when trained on only 25% of the available dataset.
Conclusion
By bridging the gap between raw computational power and human-like cognitive reasoning, this research provides a more robust framework for trajectory prediction in complex, mixed-autonomy environments. Future work will focus on optimizing computational overhead for real-time urban deployment and exploring how the model generalizes across different cultural driving norms.
Summary updated