Abstract
Previous researches reported that a lane change is the main cause of vehicle accidents. Therefore, inferring behaviors in other drivers and alarming to a driver are crucial tasks for safety. In this paper we propose a method to estimate lane change intentions in other drivers using Hidden Markov Models (HMMs). Comparing the proposed method with previous researches, we confirmed this method has higher accuracy to detect a lane change in other vehicles. And the method can recognize a slow lane change which is remained as a problem.
Introduction
Predicting the intentions of surrounding drivers is a cornerstone of autonomous vehicle safety and Advanced Driver Assistance Systems (ADAS). While detecting your own vehicle's lane-change intent is relatively straightforward due to access to internal sensors like steering wheel angle and turn signals, inferring the intent of a "target" vehicle is significantly more challenging.
Existing methods often face two primary hurdles:
- Data Accessibility: Many models rely on internal vehicle data (steering angle, velocity) that external sensors like LiDAR or cameras cannot easily measure in surrounding cars.
- The "Slow Drift" Problem: Methods utilizing the statistical distribution of a vehicle's lateral position often fail to distinguish a very slow, gradual lane change from standard "lane keeping," where a car naturally wobbles slightly within its lane. To a distribution-based model, these behaviors appear identical.
Methods
The researchers employ a Hidden Markov Model (HMM) with a left-to-right structure, specifically designed to model processes that occur in a fixed sequence.
State Definition
The model defines three distinct states to represent the phases of a lane change: Keeping (staying in the current lane), Changing (crossing the lane marker), and Adjustment (stabilizing in the target lane).
Feature Selection
The primary innovation of this paper lies in its feature selection. Instead of relying on steering data or position distributions, the authors use a vector of four lateral positions ($x_0, x_1, x_2, x_3$). These represent the vehicle's lateral position at its current location and at points 10m, 20m, and 30m directly ahead, calculated based on current lateral and longitudinal velocity.
By calculating lateral positions for points 10–30 meters ahead, the model essentially "looks into the future" of the trajectory, making the intent to change lanes much more apparent before the car even crosses the line. To ensure real-time performance, the system uses a constant centerline as a reference and implements a "limit of deviations" to prevent rapid maneuvers from destabilizing the HMM’s probability calculations.
Datasets
The model was evaluated using the Next Generation Simulation (NGSIM) dataset provided by the U.S. Federal Highway Administration. This standard real-world dataset for traffic flow and vehicle trajectories provided 100 samples of actual lane changes and 100 samples of lane keeping for the evaluation set.
Key Results
The proposed HMM method demonstrated high reliability across the board:
- Lane Change Detection: 99% accuracy (99 out of 100 samples correctly identified).
- Lane Keeping Recognition: 95% accuracy.
- Performance Comparison: The model matched the performance of systems using internal steering data—which are impractical for detecting other drivers—and significantly outperformed previous HMM and SVM models that relied on position distributions, which typically achieve only ~80% accuracy for slow changes.
- Qualitative Success: The model successfully identified "slow lane changes" where the vehicle's lateral movement was so gradual it was previously indistinguishable from normal lane-keeping noise.
Conclusion and Future Work
By restricting the model to only lateral position data—which is easily measurable via LiDAR or GPS—the authors created a system deployable in real-world traffic scenarios where "insider info" on other cars is unavailable.
Future work will look to address current limitations, such as the assumption that a vehicle's trajectory remains constant over the 30m look-ahead distance. Additionally, integrating environmental context, such as the positions of other surrounding cars or upcoming exits, could further refine the model's ability to predict driver intent in complex urban environments.
Summary updated