DNC-IMM: Early Lane-Change Intention Recognition via Neural Calibration Based on Driving Context Information

• Existing methods often detect lane changes only after lane-boundary crossing, limiting time for proactive collision avoidance, gap maintenance, and path planning.Source: §I · p. 1
• Physics-based IMM methods are interpretable but fixed transition probabilities and likelihood models cannot fully reflect rapidly changing traffic context.Source: §I · p. 1
• DNC-IMM calibrates IMM transition probabilities and measurement likelihoods using target-vehicle and surrounding-traffic features.Source: §I · p. 1
• The neural network outputs corrections rather than direct maneuver classes, so the final LK, LCL, or LCR decision comes from the calibrated IMM mode posterior.Source: §I · p. 1
• Comparative experiments show strong recognition 1–3 s before crossing, with the best average macro F1 at the more challenging 2–3 s horizons.Source: §I · p. 1

• Physics-based approaches infer maneuver intention from explicit vehicle-motion models, probabilistic state transitions, and measurements.Source: A. Physics-Based Intention Estimation · p. 2
• Adaptive methods adjust model parameters or prediction thresholds using driving context, but rule-based corrections depend on designed interaction criteria.Source: B. Rule-Based Adaptive Calibration · p. 2
• Deep neural methods learn complex trajectory and interaction patterns directly, but their reasoning may be opaque or inconsistent with vehicle dynamics and roadway constraints.Source: C. Data-Driven Learning Models · p. 2
• DNC-IMM uses a Neural Calibration Branch to learn context corrections and an IMM Estimation Branch to incorporate them into Kalman-filter-based multi-model estimation.Source: §III · p. 2

1) Input Composition and Shared Encoder:

• A shared MLP encodes target-vehicle features, surrounding context, one-second history statistics, and the prediction horizon.Source: Figure 1 · p. 2, A. Neural Calibration Branch · p. 2, 1) Input Composition and Shared Encoder · p. 2
• The branch receives a normalized 186-dimensional input vector formed from target-vehicle and surrounding-context statistics plus a prediction-horizon indicator.Source: 1) Input Composition and Shared Encoder · p. 2
• The shared encoder produces a 64-dimensional hidden representation that feeds two independently parameterized calibration heads.Source: 1) Input Composition and Shared Encoder · p. 2

• The transition head outputs a bounded 3 × 3 transition-calibration matrix, while the likelihood head outputs a 3 × 1 vector for LK, LCL, and LCR.Source: 2) Transition and Likelihood Calibration · p. 2
• Tanh scaling bounds corrections to [−3, 3], and mean removal retains relative likelihood scores because a shared constant cancels in softmax.Source: 2) Transition and Likelihood Calibration · p. 3
• The correction signals are used only to calibrate IMM transition priors and measurement likelihoods, not to directly predict maneuvers.Source: 2) Transition and Likelihood Calibration · p. 3

• The branch receives previous state estimates, covariances, and mode probabilities for each maneuver model before performing standard IMM operations.Source: B. IMM Estimation Branch · p. 3
• The resulting mode posterior µ_t provides the final maneuver decision and supports the final state and maneuver outputs.Source: B. IMM Estimation Branch · p. 3, Figure 2 · p. 3
• Learned transition and likelihood corrections enter separate stages of the standard IMM procedure.Source: Figure 2 · p. 3

State definition

• The state and observation representation jointly describes longitudinal and lateral motion using positions, velocities, and accelerations.The observation includes measurable position and velocity terms, while acceleration is inferred during Kalman estimation.Source: 1) State Definition and Mode-Wise… · p. 3

Mode-wise estimation

• The IMM simultaneously maintains LK, LCL, and LCR motion modes.Source: 1) State Definition and Mode-Wise… · p. 3
• A Kalman update is performed independently for every mode using the current observation.Source: 1) State Definition and Mode-Wise… · p. 3
• The three filters retain distinct state and covariance estimates under the same observation because they use different maneuver assumptions.Source: 1) State Definition and Mode-Wise… · p. 3
• The IMM evaluates lane keeping and both lane-change directions in parallel rather than relying on a single motion model.Source: 1) State Definition and Mode-Wise… · p. 3

Transition calibration

• The neural transition-calibration matrix is combined with the base transition matrix in log space.Source: 2) Transition and Likelihood Matrix… · p. 3
• Row-wise softmax normalization keeps the calibrated transition matrix valid by making every row sum to one.The coefficient βtrans controls the strength of transition calibration.Source: 2) Transition and Likelihood Matrix… · p. 3

Likelihood calibration

• A centered neural likelihood score calibrates each Kalman filter's measurement likelihood.The original log likelihood of mode j is used as the quantity being calibrated.Source: 2) Transition and Likelihood Matrix… · p. 3

Posterior decision

• The calibrated prior and likelihood are combined to obtain the final mode posterior, from which the maneuver estimate is selected.Source: 2) Transition and Likelihood Matrix… · p. 3
• The network cannot bypass the IMM with a direct classification output, so each decision remains traceable to calibrated mode probabilities.Source: 2) Transition and Likelihood Matrix… · p. 3

Posterior alignment

• The final IMM posterior, rather than a separate classifier output, serves as the learning target.Training encourages transition and likelihood corrections to increase the posterior probability of the ground-truth maneuver.Source: 3) Posterior-Alignment Loss · p. 3

Constraint terms

• The loss combines posterior alignment with regularization that limits the magnitude of neural corrections.The coefficient λ∆ controls the contribution of the correction-magnitude regularizer.Source: 3) Posterior-Alignment Loss · p. 3
• A lane-availability term suppresses LCL or LCR when the corresponding adjacent lane does not exist.The coefficient λinvalid determines the strength of this constraint.Source: 3) Posterior-Alignment Loss · p. 3
• The objective combines posterior alignment with safeguards against overly large corrections and physically impossible lane-change predictions.Source: 3) Posterior-Alignment Loss · p. 3

Evaluation Setup

• The evaluation measures LK, LCL, and LCR mode posteriors at 1, 2, and 3 s before lane crossing on highD.The dataset provides vehicle kinematics, lane information, and surrounding-vehicle gaps and relative velocities; recordings are split into training, validation, and testing groups.Source: §IV · p. 3, A. Experimental Setup · p. 4

Evaluation Metrics

• Macro F1 gives equal weight to LK, LCL, and LCR, making it suitable for the class imbalance in early lane-change recognition.Performance is evaluated with both class-wise F1 and macro F1.Source: B. Comparative Results · p. 4

Comparative Performance

• 0.9185 average macro F1 is obtained by DNC-IMM at the earlier 2–3 s horizons, the highest among the compared methods.DNC-IMM is slightly below Shi et al. at 1 s and in the all-horizon average, but exceeds the long-horizon averages of Mozaffari et al. and Liu et al.Source: B. Comparative Results · p. 4

Posterior Trajectories

• The true lane-change posterior exceeds 0.5 before lane crossing in representative LCL and LCR scenarios, while P(LK) remains high in the LK scenario.The trajectories show probability mass moving toward the correct lane-change direction during the developing maneuver.Source: B. Comparative Results · p. 4, Figure 3 · p. 4

Interpretability

• The calibrated posterior provides an interpretable account of probability mass moving from LK toward LCL or LCR as intention develops.This complements the final class decision with a probabilistic trajectory.Source: B. Comparative Results · p. 4

Configurations

• Four configurations are compared: DNC-IMM, Transition-only, Likelihood-only, and No Neural IMM.The first uses both corrections, the next two retain one correction, and the last removes learned calibration.Source: C. Component Ablation · p. 4, Table III · p. 4

Ablation Design

• The ablation isolates how transition calibration and likelihood calibration contribute to test-set performance.The comparison is reported as component ablation results on the test set.Source: C. Component Ablation · p. 4, Table III · p. 4

Signal and Posterior Alignment

• Figure 5 aligns adjacent-lane front gap, lateral velocity, correction signals, transition probabilities, and the final IMM posterior on one time axis.The panels connect context changes to calibrated inference in an LCR case.Source: D. Case Analysis of Neural Calibration… · pp. 4–5, Figure 5 · p. 5

Component Effect

• DNC-IMM identifies LCR from 3 s before lane crossing, whereas single-calibration variants respond later and uncalibrated IMM remains biased toward LK.The qualitative ablation highlights the joint use of transition and likelihood calibration.Source: Figure 4 · p. 5, §V · p. 5

Context-Driven Corrections

• Increasing lateral motion and changing adjacent-lane front gap coincide with increases in the LCR likelihood and LK-to-LCR transition corrections.The neural branch transforms context into both correction types.Source: D. Case Analysis of Neural Calibration… · p. 5

Posterior Shift

• The calibrated LK-to-LCR transition probability rises above its base value, followed by a posterior shift from LK to LCR before lane crossing.This traces the complete calibration-to-decision chain.Source: D. Case Analysis of Neural Calibration… · p. 5, Figure 5 · p. 5

Interpretation

• The conclusion characterizes DNC-IMM as combining data-driven context sensitivity with physically grounded probabilistic reasoning.This follows from neural calibration of transition probabilities and measurement likelihoods while retaining the IMM posterior decision.Source: §V · p. 5