Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction
1 st Jiazhao Shi*, 2 nd Qiyang Xie, 3 rd Ziyu Wang, 4 th Yichen Lin, 5 th Di Zhu, 6 th Chen Xie, 7 th Ziwei Wang, 8 th Haoyun Zhang, 9 th Enliang Li, 10 th Zetong Guan · 2025
Read Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction by 1 st Jiazhao Shi*, 2 nd Qiyang Xie, 3 rd Ziyu Wang, 4 th Yichen Lin, 5 th Di Zhu, 6 th Chen Xie, 7 th Ziwei Wang, 8 th Haoyun Zhang, 9 th Enliang Li, 10 th Zetong Guan (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-24075.txt.
Introduction
Jiazhao Shi and five colleagues predict lane changes before they happen. Driver-assistance systems need such warnings a few seconds early. A driver's intent depends on more than the car's current state. Risk builds and fades over time. Gaps open and close. The target lane may not be usable.
The framework turns these pressures into time-based descriptors. They cover risk evolution and gap persistence. They also cover the value of the other lane. Interaction pressure from nearby vehicles is included. So are maneuver feasibility and intent consistency. These signals fuse with patterns learned from raw trajectories. The combined representation classifies the outcome. Choices are left change, right change, or no change. On two German highway datasets, scores stayed high. Results were especially strong near ramps.
Catalog record
| CREATOR | 1 st Jiazhao Shi*, 2 nd Qiyang Xie, 3 rd Ziyu Wang, 4 th Yichen Lin, 5 th Di Zhu, 6 th Chen Xie, 7 th Ziwei Wang, 8 th Haoyun Zhang, 9 th Enliang Li, 10 th Zetong Guan |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-24075
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Evolutionary Physics-Informed Temporal Fusion for Lane-Change Intention Prediction 1 st Jiazhao Shi*, 2 nd Qiyang Xie, 3 rd Ziyu Wang, 4 th Yichen Lin, 5 th Di Zhu, 6 th Chen Xie, 7 th Ziwei Wang, 8 th Haoyun Zhang, 9 th Enliang Li, 10 th Zetong Guan Abstract Early lane-change intention prediction is essential for autonomous driving and ADAS, but it remains challenging because lane-changing behavior depends on evolving traffic risk, surrounding-vehicle interactions, and target-lane feasibility rather than only instantaneous vehicle states. This study proposes an evolutionary physics-informed temporal fusion framework for three-class lane-change intention prediction, including left lane change, right lane change, and no lane change. Instead of using static physics-informed variables alone, the proposed method derives temporal descriptors from conventional traffic signals, including risk evolution, gap persistence, counterfactual lane utility, interaction pressure gradient, maneuver feasibility, and intent consistency. These descriptors are fused with temporal embeddings learned from raw trajectory sequences through a sequence encoder, and the fused representation is used for final classification. Experiments are conducted on the highD and exiD datasets under 1 s, 2 s, and 3 s prediction horizons. The proposed model achieves Macro F1-scores of 0.9514, 0.9256, and 0.8872 on highD, and 0.9386, 0.9070, and 0.8531 on exiD, respectively. The improvement is especially pronounced in exiD ramp-adjacent scenarios, indicating that temporal physical evolution is particularly useful in interaction-rich environments. These results demonstrate that combining evolutionary physics-informed descriptors with learned temporal representations provides a more dynamic and interpretable solution for early lane-change intention prediction. I Introduction I-A Background and Motivation Lane-change intention prediction is a fundamental problem in autonomous driving systems and Advanced Driver Assistance Systems (ADAS), because lane-changing behavior directly affects motion planning, collision avoidance, and risk-aware decision-making. Unlike steady lane keeping, a lane-change maneuver is not determined solely by the ego vehicle’s instantaneous motion state. Instead, it emerges from a dynamic decision process involving recent ego-vehicle motion, surrounding-vehicle interactions, adjacent-lane opportunities, and the driver’s response to changing traffic constraints. Classical lane-changing theory has emphasized that lane-change decisions are governed by both incentive and safety considerations, including the expected benefit of moving to a target lane and the possible risk imposed on neighboring vehicles [ 1 , 2 ] . With the availability of high-resolution naturalistic trajectory datasets, learning-based lane-change prediction has become an active research direction. Existing studies have incorporated trajectory data, surrounding-vehicle context, and road-environment information to predict lane-change decisions or future trajectories [ 3 ] . The exiD dataset further extends highway trajectory research from regular straight-road segments to highly interactive highway entry and exit areas, providing a more challenging benchmark for modeling merging, diverging, and ramp-adjacent behaviors [ 4 ] . These developments suggest that lane-change prediction should be treated not only as a motion-pattern recognition problem, but also as a context-dependent temporal reasoning problem. Early lane-change prediction is especially challenging because useful predictions must be made before the vehicle crosses the lane boundary or fully enters the target lane. At longer prediction horizons, however, observable intention cues are often weak, incomplete, or ambiguous. Slight lateral drift, small heading variation, or gradual speed adjustment may indicate an emerging maneuver, but such cues can also be confused with normal lane-keeping fluctuations, especially in dense traffic. Therefore, early lane-change intention prediction requires models that capture not only the current traffic state, but also how maneuver-related physical signals evolve over time. Our previous study proposed a physics-informed AI framework for three-class lane-change intention prediction across multiple highway scenarios, demonstrating that domain-informed variables can improve interpretability and predictive performance in both straight-road and ramp-related environments [ 5 ] . However, that work mainly used physics-informed variables as structured representations of the traffic state. Although such features are useful, they primarily describe what the physical state is at a given moment rather than how the state evolves before a lane-change decision becomes explicit. This motivates the present study, which reformulates physics-informed lane-change prediction as a temporal evolution modeling problem. I-B Limitations of Static Physics-Informed Modeling Physics-informed variables such as Time-to-Collision (TTC), Time Headway (THW), Distance Headway (DHW), relative distance, relative velocity, lane offset, and target-lane gap size provide interpretable descriptions of motion dynamics, safety margins, and interaction feasibility. Recent lane-change risk and behavior-prediction studies continue to show the value of integrating traffic context, driving style, and interpretable risk-related variables into learning-based models [ 6 , 7 ] . However, many conventional physics-informed variables are static or only weakly temporal when used directly. A single TTC value describes the estimated collision time at one moment, but it does not indicate whether risk is worsening, recovering, or fluctuating. Similarly, a target-lane gap size may indicate that an adjacent-lane opportunity exists at the current frame, but it does not reveal whether the gap is opening, closing, or likely to remain feasible long enough for a maneuver. Therefore, the key predictive evidence for early lane-change intention may lie not only in the instantaneous value of a physical variable, but also in its temporal behavior over the observation window. This limitation is particularly important in complex traffic scenarios. In relatively regular straight-road environments, static traffic-state variables may already provide strong predictive cues. In ramp-adjacent environments, however, vehicles face changing lane topology, merging and diverging flows, and rapidly evolving interaction constraints. Recent context-aware and topology-aware lane-change studies have highlighted the importance of structured interaction modeling for complex traffic scenarios [ 7 ] . More broadly, recent spatiotemporal traffic modeling work also suggests that traffic prediction benefits from architectures that explicitly capture temporal dynamics and interaction structure rather than relying only on static representations [ 8 ] . I-C Core Idea: Evolutionary Physics-Informed Temporal Modeling This study reframes physics-informed lane-change prediction as a temporal evolution modeling problem rather than a static feature classification problem. Instead of using TTC, THW, DHW, gap size, relative motion, and lane offset only as instantaneous inputs, we derive evolutionary physics-informed descriptors that summarize how risk, opportunity, interaction pressure, and maneuver feasibility change within the observation window. Specifically, the proposed descriptor system includes six groups: risk evolution, gap persistence, counterfactual lane utility, interaction pressure gradient, maneuver feasibility, and intent consistency. These descriptors represent whether safety risk is deteriorating or recovering, whether a target-lane opportunity is stable or transient, whether moving left or right would reduce projected risk, whether surrounding-vehicle pressure changes directionally, whether the maneuver is physically feasible, and whether multiple weak cues support the same maneuver direction. In this way, physically meaningful traffic signals are reinterpreted as dynamic evidence of emerging lane-change intention rather than as fixed snapshots of the traffic state. To further exploit temporal information, this study develops a temporal-physics fusion framework for three-class lane-change intention prediction. The temporal branch learns latent maneuver representations from raw trajectory sequences, while the physics branch constructs evolutionary descriptors from physically meaningful traffic variables. The two information streams are fused and passed to a classifier to predict left lane change, right lane change, or no lane change. This design combines the representation capacity of sequence learning with the interpretability and robustness of physics-informed modeling. Recent studies on lane-change maneuver and risk prediction have shown the value of jointly modeling maneuver probability, risk level, and time-related behavior [ 9 ] , while sequence models such as LSTM-based and Transformer-based