Automatic Detection Method of Lane-Changing Intentions Based on Relationship with Adjacent Vehicles Using Artificial Potential Fields

Hanwool Woo, Yonghoon Ji, Hitoshi Kono, Yusuke Tamura, Yasuhide Kuroda, Takashi Sugano, Yasunori Yamamoto, Atsushi Yamashita, Hajime AsamaInternational Journal of Automotive Engineering 2016 · DOI

Abstract: This paper presents a new method to detect lane changes of other vehicles automatically. The main contribution of this work is to propose a new feature using a potential field that changes the distribution depending on the relative number of adjacent vehicles. Previous researches have considered only some of limited situations, for example, that the preceding vehicle is slower than the target vehicle. Therefore, degradation of the detection performance can occur under conditions that were not considered. On the other hand, the new feature we propose is able to describe general lane-changing situations by applying a dynamic potential model. We trained an estimation model and evaluated the performance using a traffic dataset with over 900 lane changes. It was confirmed that the proposed method outperforms previous methods in terms of both accuracy and early detection.

Introduction

Predicting the intentions of other drivers is a critical challenge for autonomous vehicle safety. While detecting a lane change after it has commenced is relatively straightforward, predicting the maneuver before it occurs is significantly more difficult. Most existing research relies on internal vehicle data—such as steering wheel angle or driver eye movement—which are inaccessible to external observers. Methods that do analyze external vehicles often oversimplify the environment, assuming lane changes are primarily driven by a slow preceding vehicle. In reality, drivers react to a complex "social field" of traffic, including tailgating vehicles or openings in adjacent lanes. This paper introduces a method to capture these environmental relationships using Artificial Potential Fields (APF) to provide earlier and more reliable warnings.

Methods and System Architecture

The proposed system utilizes a two-pronged approach to feature extraction, feeding into a probabilistic classifier. The overall architecture encompasses data acquisition from position sensors, lane information, and laser scanners, followed by coordinate transformation and feature extraction.

Overview of the proposed lane-change estimation system architecture, showing the flow from sensor input to state estimation.

The core components include:

  1. Driving Features: Basic metrics such as the target vehicle’s distance from the lane centerline and its lateral velocity.
  2. The Potential Feature: This "secret sauce" treats vehicles like charged particles in a physics simulation. Using a dynamic potential model based on the von Mises distribution, the system calculates a "repulsive" energy around every car.
  3. Estimation Model: These features are processed by a Hidden Markov Model (HMM) with an ergodic structure. The HMM categorizes the vehicle’s state into four intentions: keeping, changing, arrival, and adjustment.
  4. Signal Processing: To mitigate sensor noise, a dual-layer filter (Kalman Filter followed by a Moving Average Filter) is implemented.

Modeling Traffic "Social Pressure"

A key insight of this work is that a lane change is often a reaction to the surrounding traffic "field." By modeling the relationship with lead, following, preceding, and rear vehicles simultaneously, the model captures the underlying motivation for the movement.

Spatial relationship between the target vehicle and its adjacent neighbors (lead, following, preceding, and rear) used to calculate the potential field.

If the potential energy of the current lane is significantly higher than the target lane, a lane change becomes mathematically "advantageous." This allows the model to predict maneuvers even before lateral movement is detected. Furthermore, the ergodic HMM structure allows the system to handle "aborted" maneuvers, where a driver starts to change lanes but returns to their original lane.

Experimental Evaluation

The study utilized two primary data sources: the NGSIM (Next Generation Simulation) dataset, providing 958 lane-changing events from the I-80 freeway, and Experimental Vehicle Data collected in Tokyo using a sensor-equipped rig (GNSS/IMU and laser scanners).

The proposed method was compared against established baselines (Mandalia and Schlechtriemen) with the following results:

  • Accuracy: The system achieved an $F_1$ score of 97.2%, outperforming the best baseline (96.3%).
  • Early Detection: On average, lane changes were detected 1.88 seconds before the vehicle crossed the centerline.
  • False Alarm Reduction: The approach reduced the false alarm rate by nearly 25% compared to previous methods, as the potential field distinguishes between lane "wandering" and intentional movement.

Conclusion

The integration of Artificial Potential Fields with Hidden Markov Models provides a robust framework for predicting lane-changing intentions. Future work will focus on evaluating the system against larger, more diverse datasets to address edge cases and integrating the model into live Advanced Driver Assistance Systems (ADAS) for real-time validation.

Summary updated

Abstract

This paper presents a new method to detect lane changes of other vehicles automatically. The main contribution of this work is to propose a new feature using a potential field that changes the distribution depending on the relative number of adjacent vehicles. Previous researches have considered only some of limited situations, for example, that the preceding vehicle is slower than the target vehicle. Therefore, degradation of the detection performance can occur under conditions that were not considered. On the other hand, the new feature we propose is able to describe general lane-changing situations by applying a dynamic potential model. We trained an estimation model and evaluated the performance using a traffic dataset with over 900 lane changes. It was confirmed that the proposed method outperforms previous methods in terms of both accuracy and early detection.
BibTeX
@article{woo2016automatic,
  author = {Hanwool Woo and Yonghoon Ji and Hitoshi Kono and Yusuke Tamura and Yasuhide Kuroda and Takashi Sugano and Yasunori Yamamoto and Atsushi Yamashita and Hajime Asama},
  title = {{Automatic Detection Method of Lane-Changing Intentions Based on Relationship with Adjacent Vehicles Using Artificial Potential Fields}},
  journal = {International Journal of Automotive Engineering},
  year = {2016},
  doi = {10.20485/jsaeijae.7.4\_127},
}
Figure