PaperlayerView exact revision on arXiv

Source-linked AI summary

A Learning-Based Approach for Lane Departure Warning Systems with a Personalized Driver Model

Wenshuo Wang, Ding Zhao, Junqiang Xi, Wei Han

arXiv:1702.01228v1cs.LGcs.SYeess.SY

TL;DR

TLC-based warning systems can mistake drivers’ corrective behavior for an impending departure, contributing to false alarms. The paper combines a personalized driver model with online trajectory prediction to guide warnings, and reports a 3% false-warning rate with a 1 s prediction time.

  • Problem

    TLC-based warnings can be triggered even when drivers are able to return the vehicle to the lane, because these systems do not infer upcoming driver behavior.

    Source: B. Related Research · p. 1

  • Method

    The method combines a GMM-HMM personalized driver model with online prediction of lateral vehicle trajectory to infer whether drivers will return to the lane or cross its boundary.

    Source: C. Contributions · p. 2

  • Results

    3% false-warning rate with 1 s prediction time was the lowest reported for the TLC-PDM method compared with basic TLC and TLC-DSPLS.

    Source: §VII · p. 11

  • Takeaways & Limitations

    The warning strategy uses predicted trajectories to allow drivers time to return to the lane rather than warning when correction is likely.

    Source: D. TLC-Based LDP with PDM · p. 7

    The prediction algorithm treats vehicle speed and road curvature as constant, omits steering angle, and is evaluated mainly on roads with small curvature.

    Source: B. Vehicle Dynamics and Road Curvature · p. 11, §VI · p. 11

Abstract

from arXiv · show

Misunderstanding of driver correction behaviors (DCB) is the primary reason for false warnings of lane-departure-prediction systems. We propose a learning-based approach to predicting unintended lane-departure behaviors (LDB) and the chance for drivers to bring the vehicle back to the lane. First, in this approach, a personalized driver model for lane-departure and lane-keeping behavior is established by combining the Gaussian mixture model and the hidden Markov model. Second, based on this model, we develop an online model-based prediction algorithm to predict the forthcoming vehicle trajectory and judge whether the driver will demonstrate an LDB or a DCB. We also develop a warning strategy based on the model-based prediction algorithm that allows the lane-departure warning system to be acceptable for drivers according to the predicted trajectory. In addition, the naturalistic driving data of 10 drivers is collected through the University of Michigan Safety Pilot Model Deployment program to train the personalized driver model and validate this approach. We compare the proposed method with a basic time-to-lane-crossing (TLC) method and a TLC-directional sequence of piecewise lateral slopes (TLC-DSPLS) method. The results show that the proposed approach can reduce the false-warning rate to 3.07\%.

I. INTRODUCTION

Lane-departure warning systems risk false alarms when they cannot distinguish unintended departures from drivers correcting their course. The paper proposes a personalized model and trajectory prediction to infer upcoming driver behavior and guide warnings.

A. Motivation

  • Lane departure is an unintentional drift toward a lane boundary, and lane-departure warning systems alert drivers when departure begins or is predicted.The introduction reports that single-vehicle lane departure accounts for 37.4% of fatal crashes in the United States.Source: A. Motivation · p. 1

B. Related Research

  • False alarms arise when TLC-based warnings fail to distinguish unintended lane departure from a driver's correction back toward lane center.Such warnings may annoy drivers, while prior approaches described here do not predict both future vehicle trajectories and driver operations.Source: B. Related Research · pp. 1–2

C. Contributions

  • The proposed approach combines a personalized driver model with online lateral-trajectory prediction to infer whether drivers will return to the lane or cross its boundary.The model represents lane-keeping and lane-departure behavior using a GMM and HMM.Source: C. Contributions · p. 2

D. Paper Organization

  • The paper is organized into sections on prediction and warnings, the proposed approach, experiments, further analysis, discussion, and conclusions.The proposed approach and model-based prediction algorithm are described in Section III.Source: D. Paper Organization · p. 2

II. LANE DEPARTURE PREDICTION

  • Lane-departure prediction estimates whether a vehicle will leave its lane, giving the driver more time to act to avoid a crash.The paper distinguishes TLC-based prediction, vehicle-variable-based position estimation, and lane-boundary detection from road images.Source: §II · p. 2

A. Time to Lane Crossing (TLC)

  • TLC-based systems estimate lane and vehicle states, calculate time to lane crossing, and alert the driver when it reaches a threshold.For small road curvature, TLC can be calculated using lateral distance and lateral velocity or distance to the line crossing.Source: A. Time to Lane Crossing (TLC) · p. 2

B. Excessive False Warning

TLC-based warnings can be false when drivers correct their course after approaching a lane boundary. The paper models driver behavior using five driving variables and combines GMM and HMM to support prediction.

B. Excessive False Warning

  • TLC-based methods can falsely warn when drivers approach a lane boundary but then steer back, a behavior that better sensor data or TLC calculations alone may not prevent.The paper frames the prediction question as whether the driver will return to lane center within a short time.Source: B. Excessive False Warning · pp. 2–3

III. PROPOSED METHOD

  • The method uses GMM components as HMM hidden modes, then uses observations to infer driver states and predict upcoming behavior.The GMM models dependencies among variables describing driver behavior.Source: §III · p. 3

A. Feature Parameter Selection

  • Five variables describe driver behavior: vehicle speed, relative yaw angle and rate, road curvature, and relative lateral displacement.The study focuses on roads with curvature ρ ≤ 10^-4 m^-1.Source: A. Feature Parameter Selection · pp. 3–4

B. GMM

The GMM models dependencies among five variables as a weighted sum of multivariate Gaussian distributions, then estimates its parameters by maximizing likelihood with an iterative EM procedure.

1) Structure of GMM:

  • The GMM represents dependencies among five variables as a weighted sum of multivariate Gaussian distributions.Each component has a weight, mean, and covariance; BIC is used to select the number of components.Source: 1) Structure of GMM · p. 4

2) Parameter Identification:

  • Maximum likelihood estimates the GMM parameters by maximizing the likelihood of the observed data sequence.Because direct optimization is difficult, the method uses iterative EM, whose likelihood increases monotonically at each step.Source: 2) Parameter Identification · p. 4
  • EM repeats its E-step, M-step, and log-likelihood update until the convergence condition is met, using ε = 10^-10.The E-step computes component posterior probabilities; the M-step updates model parameters.Source: 2) Parameter Identification · p. 4

C. HMM

The connected HMM representation builds on the trained GMM, treating each GMM component as a hidden mode for modeling driver lane-keeping behavior.

C. HMM

  • Each trained GMM component becomes a hidden mode in the connected HMM representation of driver lane-keeping behavior.The transition matrix contains probabilities of moving between modes and is estimated from training data.Source: C. HMM · p. 4
  • The HMM observation combines vehicle speed, relative yaw angle, road curvature, and lateral displacement, while relative yaw rate is treated as hidden.The hidden mode is selected from K possible modes.Source: C. HMM · p. 4
  • During inference, the model estimates relative yaw rate from the observed driving-situation sequence using a conditional expectation.The hidden-mode and observation sequence distribution is modeled using mode-conditioned multivariate Gaussian distributions.Source: C. HMM · pp. 4–5

D. Iterative Algorithm for Prediction

The iterative algorithm propagates the driver model and vehicle kinematics to predict future lateral position over q steps, under assumptions that simplify short-horizon prediction.

D. Iterative Algorithm for Prediction

  • The algorithm iteratively applies the driver model and propagates the driving situation to predict future vehicle lateral position.It uses a kinematic point-mass vehicle model and returns estimated relative lateral position at the end of the q-step horizon.Source: D. Iterative Algorithm for Prediction · p. 5
  • During prediction, vehicle speed is held constant, and emergent braking behaviors such as collision avoidance are excluded.The constant-speed assumption applies over the short prediction period.Source: D. Iterative Algorithm for Prediction · p. 5
  • Prediction assumes road curvature is constant and limits scenarios to curvature below 10^-4 m^-1; the model uses Δt = 0.1 s.The curvature assumption reduces calculation complexity.Source: D. Iterative Algorithm for Prediction · p. 5

A. Data Collection

The study collected naturalistic driving data from ten drivers, extracted lane-boundary-related events, and trained personalized models using cross-validation and BIC-based component selection.

A. Data Collection

  • Naturalistic data from ten volunteers’ own vehicles were collected through the Safety Pilot Model Deployment program using Mobileye, DSRC, and onboard data acquisition equipment.Vehicle variables came from the CAN-Bus, road variables from Mobileye, and driving data were recorded at 10 Hz during unrestricted daily trips.Source: A. Data Collection · p. 5, Figure 5 · p. 6

B. Data Preprocessing

  • Training events included DCB and/or LDB data near a road boundary, with a 15-second window before and after the case point.Case points were identified at lateral position within 0.5 m of the boundary or after boundary crossing.Source: B. Data Preprocessing · p. 6
  • Preprocessing excluded high-curvature roads, nonstandard road widths, lane changes, and events shorter than 15 seconds.Lane changes were filtered using turning-light status and lateral displacement toward an adjacent lane center.Source: B. Data Preprocessing · p. 6

C. Training Model

  • Ten-fold cross-validation trained each driver model on nine folds; BIC convergence near K = 10 set the GMM component count.The training-model figures show BIC across component counts and prediction errors and standard deviations across prediction steps.Source: C. Training Model · p. 6, Figure 6 · p. 6, Figure 7 · p. 7

D. TLC-Based LDP with PDM

The TLC-PDM warning strategy combines a TLC condition with predicted future lateral position to account for whether a driver is likely to correct course. The paper compares it with the basic TLC and TLC-DSPLS methods.

D. TLC-Based LDP with PDM

  • Warnings require a valid TLC condition plus predicted trajectory criteria that assess lane-boundary crossing and whether the driver will return toward lane center.The predicted interval is [t : t + q∆t], and the strategy is intended to give drivers time to correct without an unnecessary warning.Source: D. TLC-Based LDP with PDM · p. 7
  • The TLC threshold is set to τ = 1.0 s, while q, γ1, and γ2 determine the prediction interval and warning criteria.The authors state that a time margin greater than 0.9 s allows enough time for a driver response.Source: D. TLC-Based LDP with PDM · p. 7

E. Comparison with Other Approaches

1) Basic TLC:

  • The basic TLC method sends an alarm when condition (16) is valid.The passage characterizes this method as relatively efficient in simple driving scenarios.Source: 1) Basic TLC · p. 7

2) TLC-DSPLS:

The TLC-DSPLS baseline uses a probabilistic driver model, while the results section evaluates lateral-position prediction and false-alarm rate. The proposed TLC-PDM method predicts trajectory and whether a driver may correct a lane departure.

2) TLC-DSPLS:

  • TLC-DSPLS describes driving-signal trajectories with directional sequences of piecewise lateral slopes and uses Bayes rule to estimate LCB and DCB probabilities.The paper reproduces and re-examines this method rather than discussing it further.Source: 2) TLC-DSPLS · pp. 7–8

V. RESULTS AND ANALYSIS

  • The study analyzes lateral-position prediction performance and the false-alarm rate of the lane-departure warning system for driver correction behavior.Fig. 9 provides an example of predicted lateral position at different prediction steps.Source: §V · p. 8, Figure 9 · p. 8

A. Prediction of Lateral Positions

  • For q=5 (0.5 s), prediction error ranged from 0.063 to 0.1696 across drivers; for q=30 (3.0 s), it ranged from 0.2090 to 0.5138.Prediction performance decreases as the prediction step increases; the performance criterion uses errors between predicted and experimental trajectories.Source: A. Prediction of Lateral Positions · p. 8

B. Performances for LDW system

  • The basic TLC method can warn when its TLC condition is met without predicting whether the driver will return to lane center, whereas TLC-PDM predicts the upcoming trajectory and can withhold a warning for a slight crossing the driver is expected to correct.In the example, slight boundary crossing is less than 0.05 m; TLC-PDM warning criteria identify predicted crossing and failure to return in time.Source: B. Performances for LDW system · pp. 8–9, Figure 10 · p. 9

1) Influences of parameters q, γ1, γ2

The TLC-PDM warning frequency depends on its parameters and varies across drivers with different driving styles. The comparison evaluates false-alarm rates for TLC-PDM, basic TLC, and TLC-DSPLS.

1) Influences of parameters q, γ1, γ2

  • For the same driver, TLC-PDM warning frequency cannot exceed basic TLC; increasing γ1 or γ2 raises its frequency for fixed q.For driver #10, frequency approached the basic TLC value of 0.2034 at (γ1, γ2) = (0.6, 0.6), while remaining lower for all q.Source: 1) Influences of parameters q, γ1, γ2 · p. 9
  • With fixed q, γ1, and γ2, warning frequency differs across drivers, with centered driving associated with lower frequency and boundary-proximate driving with higher frequency.The text reports about 0.013 for drivers #5 and #6, about 0.08 for drivers #1 and #10, and greater than 0.16 for the other drivers.Source: 1) Influences of parameters q, γ1, γ2 · p. 10, Figure 12 · p. 10, Table II · p. 10

2) False-Alarm Rate:

  • The TLC-PDM false-alarm rate was 1.13% to 74.51%, compared with 18.8% to 492.2% for basic TLC and 6.44% to 198.1% for TLC-DSPLS.The authors define a DCB for this comparison as future lateral displacement greater than 0.1; they also report basic TLC exceeding 100% FAR at 1.5 s and TLC-DSPLS nearly reaching 100% at 2.0 s.Source: Figure 13 · p. 10, 2) False-Alarm Rate · p. 10

VI. FURTHER DISCUSSION AND FUTURE WORK

The discussion presents TLC-PDM as a correction-aware lane-departure warning approach, outlines limits in parameter selection and prediction scope, and reports its comparison with TLC baselines.

VI. FURTHER DISCUSSION AND FUTURE WORK

  • TLC-PDM is proposed to reduce false warnings caused by driver correction behavior.The paper also develops an online model-based prediction algorithm to forecast vehicle trajectory.Source: §VI · pp. 10–11
  • The model-based trajectory prediction showed good performance, but the paper’s framework and validation focus on driving scenarios with small road curvature.The proposed algorithm predicts vehicle trajectory and is also used as an LDW warning strategy.Source: §VI · p. 11

A. Parameter Selection

  • Warning-strategy parameters were set subjectively, but can be tuned to design priorities; larger γ1 or γ2 values prioritize safety over warning frequency.The authors also report that driving styles are highly related to warning performance.Source: A. Parameter Selection · p. 11

B. Vehicle Dynamics and Road Curvature

  • The prediction algorithm omits vehicle dynamics, holds speed and road curvature constant during iteration, and uses a simple TLC-based risk function.Future work is to consider vehicle dynamics and more complicated road profiles.Source: B. Vehicle Dynamics and Road Curvature · p. 11

C. Driver State and Intention

  • The method does not recognize drivers’ physical or physiological states or lane-change intentions; future applications should recognize lane changes.The listed states include drowsiness, fatigue, drunkenness, and aggression.Source: C. Driver State and Intention · p. 11

VII. CONCLUSION

  • TLC-PDM models five driving variables with a personalized Gaussian mixture and hidden Markov model, then predicts upcoming lateral trajectory and supports a warning strategy.The five variables are vehicle speed, relative yaw angle and its change rate, road curvature, and relative lateral displacement.Source: §VII · p. 11
  • 3% false-warning rate at 1 s prediction time was lowest against basic TLC and TLC-DSPLS.The conclusion reports validation using naturalistic driving data from 10 drivers.Source: §VII · p. 11

APPENDIX A

Appendix A describes estimating an HMM transfer matrix from training data by assigning each data point a mode and counting transitions between modes.

APPENDIX A

  • The appendix uses a training set of n data points, ξ1 through ξn, to calculate the HMM transfer matrix.It states that the total number of training points exceeds 5×10^5.Source: Appendix A · p. 11
  • Each training point is assigned a mode from the set M = {1, 2, …, K}, producing a mode sequence with one entry per data point.Source: Appendix A · p. 11
  • For each pair of modes, the transfer probability is estimated from the count of transitions from mode i to j and the total training points at mode i.The resulting probabilities form transfer matrix T.Source: Appendix A · p. 11
Loading 1702.01228v1…