Interaction-Aware Lane-Changing Early Warning

• Interaction-Aware Lane-Changing Early Warning (LCEW) systems are predictive frameworks that alert drivers before conflicts by modeling both the ego vehicle’s intent and the interactions with surrounding traffic.
• They integrate diverse methodologies such as TTC thresholding, vision-based context analysis, and probabilistic risk monitoring to anticipate lane-change events with actionable timing.
• Empirical studies and simulation evaluations show that optimizing warning timing based on real-time maneuver prediction improves situational awareness and reduces the likelihood of collision.

Searching arXiv for recent and relevant papers on interaction-aware lane-changing early warning, lane-change intention prediction, and warning generation. Interaction-Aware Lane-Changing Early Warning (LCEW) denotes a class of warning systems that attempt to alert a driver or vehicle controller before a lane-change conflict becomes immediate by combining early maneuver anticipation with some representation of interaction among the ego vehicle, surrounding traffic, the human driver, or the automation stack. In recent work, the topic spans ego-intention-conditioned TTC warning logic (Tanshi et al., 2024), human–automation interaction consistency models for suppressing inappropriate lane-departure intervention (Yan et al., 2020), vision-based anticipation of surrounding-vehicle lane changes from context-rich ROIs (Biparva et al., 2021), probabilistic feasibility monitoring for mandatory lane changes near lane drops (Mehr et al., 2020), sequential warning optimization over a 5-second horizon via a POMDP (Li et al., 2024), and fully multi-vehicle interaction-aware warning pipelines based on graph trajectory prediction and collision detection (Zhang et al., 23 Sep 2025). This suggests that LCEW is not a single algorithmic template but a family of predictive warning formulations with different notions of “interaction awareness.”

1. Conceptual scope and taxonomy

The literature distinguishes recognition from anticipation in a way that is directly relevant to early warning. In surrounding-vehicle studies based on PREVENTION, lane-change classification is defined at $TTE=0$, whereas lane-change prediction uses $TTE>0$, so the observation window ends before the event and the model must anticipate a future LLC or RLC. The event itself is defined as the frame when “the center of the rear bumper is just above the lane markings,” which makes warning lead time physically interpretable (Biparva et al., 2021). In the most explicit recent LCEW formulation, lane changing in congestion is treated as a “stochastic, multi-vehicle interactive process,” and risk is partitioned into direct and indirect components rather than reduced to a single ego-versus-nearest-neighbor conflict (Zhang et al., 23 Sep 2025).

The field can be organized into several recurrent design strands.

Strand	Representative mechanism	Interaction scope
Maneuver-conditioned TTC warning	Predicted $LK/LCL/LCR$ gates TTC thresholds (Tanshi et al., 2024)	Ego intention + current neighboring TTCs
Human–automation consistency warning suppression	GRU boundary-crossing prediction + pseudo-work consistency (Yan et al., 2020)	Driver–system interaction
Vision-based surrounding-vehicle anticipation	Two-stream, ST, and SlowFast ROI models (Biparva et al., 2021)	Implicit visual context
Probabilistic mandatory-LC warning	Trigger when $P(S)$ falls below $p_l$ (Mehr et al., 2020)	Lane-wise headway opportunity structure
Multi-vehicle predictive LCEW	STGCNN-MI + OBB collision detection (Zhang et al., 23 Sep 2025)	Direct and indirect multi-vehicle interactions

A persistent conceptual distinction runs through these strands. Some systems are ego-intention-aware but not fully interaction-aware; some are human–automation interaction-aware but weak on surrounding-traffic prediction; some are implicitly interaction-aware because they expose a network to contextual visual evidence; and some are explicitly multi-agent, using graph or game-theoretic representations of surrounding traffic. A plausible implication is that “interaction-aware” should be reserved for systems whose warning logic depends on modeled future coupling among agents, not merely on current adjacency.

2. State representation and maneuver anticipation

A concrete ego-intention-aware formulation is given by the three-class classifier over lane keeping (LK), lane change left (LCL), and lane change right (LCR) in (Tanshi et al., 2024). Its input space comprises 24 variables on a three-lane highway: ego features $Ln$, $I$, $G$, $v_{ego}$, $S$, and $TTE>0$0; and surrounding-vehicle variables for six relative positions $TTE>0$1, each with velocity, distance, and TTC. The intention recognizer is an individualized fuzzy-random forest: continuous-valued signals are fuzzified by trapezoidal membership functions parameterized by $TTE>0$2, those parameters are generated automatically via FN-DBSCAN-based clustering, and the resulting fuzzy representation is classified by RF voting (Tanshi et al., 2024). The labeling horizon is defined from $TTE>0$3 to $TTE>0$4, with $TTE>0$5 used as a proxy when the indicator is absent; the prior referenced work reported intention recognition accuracy above $TTE>0$6 for $TTE>0$7, though the same source also notes the absence of confusion matrices, class-wise precision/recall, or false-positive analysis in the current paper (Tanshi et al., 2024).

A different anticipation target appears in the haptic shared-steering literature. There, the predictor $TTE>0$8 estimates whether the ego vehicle will cross a lane boundary within the next $TTE>0$9 seconds, using a 3-second, 60 Hz temporal window of six channels: head yaw angle, longitudinal acceleration, longitudinal velocity, steering wheel angle, lateral distance to the adjacent lane, and yaw angle (Yan et al., 2020). The deployed model is a GRU-based RNN with 180 time steps and dense layers of size 128 and 128, and the reported test accuracy is $LK/LCL/LCR$0 for predicting whether the driver intends to change lane (Yan et al., 2020). For LCEW, the key point is that the anticipation target is not a static maneuver label but a near-term future lane-boundary crossing event, which is directly aligned with warning timing.

Surrounding-vehicle anticipation uses yet another representation regime. In PREVENTION-based video models, the target behavior is a 3-class LLC/RLC/NLC label for a surrounding vehicle observed from an ego forward-view camera, with ROI scales from $LK/LCL/LCR$1 to $LK/LCL/LCR$2, observation horizons of 20, 30, and 40 frames, and prediction horizons $LK/LCL/LCR$3 and $LK/LCL/LCR$4 frames, i.e. 1 s and 2 s at 10 Hz (Biparva et al., 2021). The temporal stream uses dense optical flow computed with Farnebäck polynomial expansion, and because the target is centered in the ROI, the flow primarily captures the movement of context around the target rather than simple target translation (Biparva et al., 2021). This suggests that lane-change precursors can be extracted not only from the target body but also from lane structure, neighboring vehicles, turn indicators, brake lights, and local occupancy patterns.

3. Warning-generation logics

The most direct LCEW warning rule in the literature is the maneuver-conditioned TTC gate in (Tanshi et al., 2024). The operative logic can be written as

$LK/LCL/LCR$5

The warning thresholds are maneuver- and direction-specific. For LK: $LK/LCL/LCR$6, $LK/LCL/LCR$7, $LK/LCL/LCR$8, $LK/LCL/LCR$9. For LCL: $P(S)$0, $P(S)$1, $P(S)$2, $P(S)$3. For LCR: $P(S)$4, $P(S)$5, $P(S)$6, $P(S)$7. Approval imagery uses warning thresholds plus $P(S)$8, and warnings or approvals are displayed for $P(S)$9 through directional HUD imagery plus the auditory phrase “collision risk” (Tanshi et al., 2024). The novelty is not generic collision warning but threshold switching conditioned on predicted maneuver class.

A second warning paradigm is probabilistic feasibility monitoring for mandatory lane changes near lane drops. The core quantity is

$p_l$0

the probability of successfully reaching the target lane before the lane end under current traffic conditions (Mehr et al., 2020). The online trigger is explicit: advise the driver or autonomous vehicle to start the lane-changing maneuver when $p_l$1 drops below a fixed threshold $p_l$2, with tested values $p_l$3 (Mehr et al., 2020). The model uses lane-average speeds, log-normal headway distributions, critical gaps $p_l$4 with $p_l$5 and $p_l$6, and lane-change duration $p_l$7; a microscopic safety check then verifies lead and lag gaps before execution (Mehr et al., 2020). This converts remaining opportunity, rather than raw distance-to-lane-end, into a warning signal.

A third formulation recasts warning generation itself as sequential decision-making. In the POMDP framework of (Li et al., 2024), the warning action $p_l$8 alters the latent driver policy via $p_l$9, the driver action is sampled from the post-warning policy, surrounding vehicles evolve under $Ln$0, and the planner optimizes

$Ln$1

with $Ln$2, step time $Ln$3, and receding-horizon replanning every $Ln$4, i.e. a 5-second future horizon (Li et al., 2024). This is not lane-change-intent estimation per se; one experimental scenario is a dangerous vehicle cutting into the ego lane at $Ln$5. Its relevance to LCEW lies in modeling warning-dependent driver response, delayed switching, and warning costs rather than treating warnings as one-shot threshold events.

4. Explicit interaction-aware models

The most direct recent realization of interaction-aware LCEW is the STGCNN-MI pipeline in congested traffic (Zhang et al., 23 Sep 2025). Historical trajectories of all vehicles in the lane-change ROI are represented as a graph sequence $Ln$6, with mutual-information-based edge weights

$Ln$7

used to build $Ln$8, and the graph encoder plus TXP-CNN decoder predicts future trajectories $Ln$9 for all vehicles in the ROI (Zhang et al., 23 Sep 2025). The warning module then applies OBB collision detection over predicted trajectories and issues a warning if any collision is detected within the prediction window, reporting front or rear collision location. The ROI is forward-bounded by $I$0 and rear-bounded by 250 m, and the reported empirical risk composition is 56.2% direct, 36.7% forward indirect, and 7.1% rear indirect risk (Zhang et al., 23 Sep 2025). This makes indirect risk a substantive part of the warning problem rather than a negligible side effect.

A complementary line of work extracts interpretable interaction structure from trajectory data. In the V2V primitive-based framework, a three-vehicle lane-change scenario is represented by ego, the preceding vehicle in the original lane ($I$1), and a vehicle in the target lane ($I$2); GMM-HMM segmentation decomposes 578 events into 1224 primitives, DTW-based K-means clusters them into 13 interactive patterns, and TTC-based analysis identifies cluster #12 and cluster #10 as high-risk, with mean TTCs of 7.89 s and 8.39 s, respectively (Zhang et al., 2021). The paper’s account of cluster #10 as a beginning-stage lane change in which ego is faster than $I$3, $I$4 is near the rear in the target lane, and ego starts changing lanes while reducing longitudinal speed to wait for an opportunity is especially relevant to LCEW, because it describes a pre-danger interaction configuration rather than a completed conflict.

Cooperative connected-vehicle formulations contribute a different kind of interaction awareness. The Driver Messenger System identifies the closest trailing vehicle in the adjacent target lane, using the host vehicle’s Path History rather than raw lateral distance, and sends a Driver Intent Message to that target vehicle (Shah et al., 2022). The parameter $I$5 is tested at 50, 75, 100, and 150 m; PH-based recognition outperforms lateral-distance-only recognition; and the system acts as an advanced warning mechanism by increasing time and space headway after the DIM-triggered response (Shah et al., 2022). Here interaction awareness is selective and topology-aware: the problem is not only whether a lane change is intended, but which nearby vehicle is the relevant conflict partner.

A behavioral interaction model appears in mixed-traffic game-theoretic analysis. There the lane-changing active vehicle and the target-lane lag passive vehicle are treated as players with actions $I$6, cooperative or defective. K-means clustering on 7,636 observed lane-changing events yields cooperative proportions of 69.0% for active AVs versus 57.9% for active HDVs, and 33.1% for passive AVs versus 26.1% for passive HDVs; QRE then gives cooperation probabilities

$I$7

from pre-lane-change state variables including active speed statistics, lead gap, lag gap, and relative speeds (Chung et al., 8 Dec 2025). The same framework finds social dilemmas in approximately 4% of events from the active perspective and 11% from the passive perspective, with most classified as Stag Hunt or Prisoner’s Dilemma (Chung et al., 8 Dec 2025). This suggests that some LCEW failures may arise from strategic miscoordination rather than pure kinematic insufficiency.

5. Empirical evidence and human factors

The intention-conditioned warning system in (Tanshi et al., 2024) was evaluated in a fixed-base SCANeR simulator with 270° field of view, 20 Hz acquisition, and 44 licensed drivers divided into 22 experimental and 22 control participants. The paper reports significantly improved performance ($I$8) for front vehicles during LCL and for front-right and back-right vehicles during LCR, but significantly worse performance ($I$9) for back and back-left vehicles during LCL and front vehicles during LCR; it also states that the system improved performance for lane changes left and right, but not lane keeping (Tanshi et al., 2024). Questionnaire responses were likewise mixed: at least 75% rated the system helpful, situational-awareness-increasing, and timely, more than 70% found it desirable for everyday use, but more than 50% found lane-change approval imagery not useful because it appeared too late, and about 30% reported ignoring some warnings because they themselves were anticipating surrounding vehicles’ future behavior in ways the current TTC rule did not capture (Tanshi et al., 2024). The HMI lesson is precise: warning timing benefited from intention gating, whereas “safe to go” approval required tighter alignment with the decision phase.

Video-based surrounding-vehicle anticipation reports high overall accuracy but exposes an important evaluation caveat. On PREVENTION, the best 2-second anticipation result is 91.94% overall accuracy for the Spatiotemporal Multiplier Network with ROI $G$0 and $G$1 frames, yet the corresponding confusion matrix gives LLC precision 63.5%, RLC precision 72.5%, NLC precision 97.7%, and recalls 71.7%, 74.6%, and 96.2%, respectively (Biparva et al., 2021). The implication is that headline accuracy can be dominated by the NLC majority class; for LCEW, positive-class precision and recall are operationally more important than aggregate accuracy.

Human performance provides a stringent baseline. In the PREVENTION human study, 72 participants watched front-camera clips and pressed a key when they believed a target vehicle was about to change lanes. The mean delay relative to lane-change start $G$2 was $G$3 with $G$4, the mean anticipation relative to the lane-change middle point $G$5 was $G$6 with $G$7, and the null hypothesis of zero delay was rejected with $G$8 for the two-sided t-test and $G$9 for the one-sided test (Quintanar et al., 2020). NLC clips produced 1.85 errors per user on average, with $v_{ego}$0, and the delay–accuracy relationship was summarized by $v_{ego}$1 (Quintanar et al., 2020). The central empirical point is that most participants detected lane changes after they had started, and earlier responses increased error. LCEW is therefore not merely an automation analogue of ordinary human scene reading; it seeks to surpass human onset detection by exploiting weak temporal and interaction cues systematically.

Human–automation interaction studies support the same caution about nuisance intervention. In the shared-steering experiment, SDLP differed significantly across manual, strong, and weak groups, $v_{ego}$2, $v_{ego}$3, and both strong and weak IBHS reduced SDLP versus manual; the paper also reports that whenever inconsistency was detected, the system adjusted correctly and followed the driver (Yan et al., 2020). This supports the use of interaction-consistency logic as a warning suppressor or validator when lane departure is predicted but deliberate lane change remains plausible.

6. Open issues and likely trajectories

Several limitations recur across the literature. The maneuver-conditioned TTC system in (Tanshi et al., 2024) is explicitly strongest as evidence that intention-aware warning timing matters, but it is “partially interaction-aware” in sensing and “not fully interaction-aware” in prediction, because surrounding vehicles are modeled via current distances, velocities, and TTC rather than future intentions or trajectories. The shared-steering GRU and pseudo-work model is rich in driver–automation interaction cues but “weak on traffic interaction awareness” because it has no gap-acceptance model, surrounding-vehicle trajectory prediction, or adjacent-lane safety criterion (Yan et al., 2020). PREVENTION-based video models are interaction-aware only implicitly through ROI context, with no graph relations, social pooling, lane occupancy maps, or explicit pairwise reasoning (Biparva et al., 2021). Cooperative messaging improves target selection and headway but depends on connectivity and does not provide a full collision-prediction model (Shah et al., 2022).

Recent work indicates one plausible convergence path. A hybrid multi-scenario predictor combines a two-layer Bi-LSTM encoder with kinematics-, safety-, and interaction-aware features such as headway, TTC, and safe-gap indicators, then classifies No-LC, Left-LC, and Right-LC with LightGBM. On location-based splits, it reports macro-F1 of 0.9562, 0.9124, and 0.8345 on highD and 0.9247, 0.8197, and 0.7605 on exiD at $v_{ego}$4, respectively (Shi et al., 30 Dec 2025). Yet that work is still an intention-prediction backbone rather than a complete warning policy. A plausible implication is that mature LCEW will combine four layers that are often separated in current papers: explicit maneuver anticipation, explicit surrounding-agent prediction, closed-loop warning generation that models driver response, and calibrated warning thresholds tied to class imbalance and nuisance-alarm control.

The more ambitious formulations already point in this direction. The POMDP warning generator explicitly models warning-dependent driver behavior and surrounding-agent evolution over a multi-step horizon (Li et al., 2024), while the graph-based congested-traffic LCEW explicitly predicts all-vehicle trajectories and checks direct and indirect collisions (Zhang et al., 23 Sep 2025). What remains incomplete is the unification of these pieces: uncertainty-aware multi-agent forecasting, behavior-conditioned warning optimization, and human factors that distinguish helpful advance warning from over-warning or misplaced “approval.” In that sense, LCEW is moving from thresholded local conflict indication toward a predictive, multi-agent, human-in-the-loop decision problem whose central object is not the lane change alone, but the evolving interaction field around it.