Interacting Multiple Model Vehicle-Tracking System Based on Neural Network

Jae-Pil Hwang, Seong-Keun Park, Euntai KimJournal of Korean Institute of Intelligent Systems 2009 · DOI

Abstract

본 논문에서는 새로운 방식의 적응형 순항제어 필터링 방식을 제안한다. 제안한 알고리즘은 선행 차량의 모드를 추정하는 문제를 분류기의 문제로 보고 신경망 분류기를 이용하여 이를 수행한다. 신경망은 각 모드에 대한 사후 확률을 출력하며 이를 IMM과 결합하여 선행차량의 추적을 수행한다. 끝으로 10가지 시나리오에 대하여 신경망 분류기와 IMM을 결합한 NIMM (Neural Network IMM)을 적용하여 성능을 확인한다.

In this paper, a new filtering scheme for adaptive cruise control (ACC) system is presented. In the proposed scheme, the identification of the mode of the preceding vehicle is considered as a classification problem and it is done by a neural network classifier. The neural network classifier outputs a posterior probability of the mode of the preceding vehicle and the probability is directly used in the IMM framework. Finally, ten scenarios are made and the proposed NIMM is tested on them to show its validity.

Introduction

Adaptive Cruise Control (ACC) systems require precise tracking of preceding vehicles to ensure safety and comfort. A significant challenge in vehicle tracking is the non-linear and multi-modal nature of driving behavior, where a vehicle may switch between distinct states such as constant velocity cruising or lateral lane changes. The Interacting Multiple Model (IMM) filter is a widely used framework for such problems, employing parallel Kalman filters to handle different motion models. However, traditional IMM filters rely on a fixed Markovian transition matrix to estimate mode probabilities, which often fails to capture the intentional patterns of human driving, leading to detection lag and tracking errors.

Proposed Method: Neural Network IMM (NIMM)

This paper introduces the Neural Network Interacting Multiple Model (NIMM) algorithm. The core innovation lies in replacing the static Markovian transition probabilities with a dynamic Neural Network (NN) classifier.

The Interacting Multiple Model (IMM) framework components: Interaction (A), Filtering (B), Mode Probability Update (C), and Combination (D).

The NIMM architecture integrates a multi-layer perceptron into the IMM loop. The NN is trained to recognize driving patterns from a window of historical sensor data, including position, velocity, and acceleration. A critical theoretical contribution is the demonstration that the NN's output approximates the posterior probability of the vehicle's mode. This allows the NN output to be directly utilized in the "Mode Probability Update" stage (Block C in Figure 1), enabling the system to anticipate maneuvers based on recognized trajectories rather than reacting solely to statistical residuals.

Experimental Setup

The proposed NIMM was evaluated using ten simulated driving scenarios involving "cut-in" and "cut-out" maneuvers at high speeds (30–33 m/s). To ensure robustness, a 10-fold cross-validation approach was employed, training the NN on nine scenarios and testing on the tenth. This methodology prevents the model from over-fitting to specific lane-change geometries and tests its generalization to varied driving patterns.

Results and Discussion

The performance of NIMM was compared against a standard IMM filter. Quantitative results show that NIMM achieved a lower Mean Squared Error (MSE) of 0.525 compared to 0.542 for the standard IMM.

Simulation results for various lane-change scenarios (a-f), showing the tracking performance of the proposed system.

Qualitative analysis of the tracking plots reveals that NIMM identifies mode transitions with significantly less delay. By recognizing the pattern of a lane change early in its execution, the NIMM-based tracker aligns more closely with the actual vehicle trajectory, reducing the "overshoot" and lag typically seen in reactive Markovian filters. The neural network's ability to output posterior probabilities directly into the IMM framework ensures that the most appropriate motion model is prioritized during critical maneuvers.

Conclusion

The NIMM algorithm successfully bridges the gap between statistical filtering and pattern recognition. By treating mode transitions as predictable maneuvers rather than random events, the system provides more accurate and responsive tracking for ACC applications. Future work will focus on testing the system with real-world radar data and optimizing the computational efficiency for automotive hardware.

Summary updated

Abstract

본 논문에서는 새로운 방식의 적응형 순항제어 필터링 방식을 제안한다. 제안한 알고리즘은 선행 차량의 모드를 추정하는 문제를 분류기의 문제로 보고 신경망 분류기를 이용하여 이를 수행한다. 신경망은 각 모드에 대한 사후 확률을 출력하며 이를 IMM과 결합하여 선행차량의 추적을 수행한다. 끝으로 10가지 시나리오에 대하여 신경망 분류기와 IMM을 결합한 NIMM (Neural Network IMM)을 적용하여 성능을 확인한다. In this paper, a new filtering scheme for adaptive cruise control (ACC) system is presented. In the proposed scheme, the identification of the mode of the preceding vehicle is considered as a classification problem and it is done by a neural network classifier. The neural network classifier outputs a posterior probability of the mode of the preceding vehicle and the probability is directly used in the IMM framework. Finally, ten scenarios are made and the proposed NIMM is tested on them to show its validity.
BibTeX
@article{hwang2009interacting,
  author = {Jae-Pil Hwang and Seong-Keun Park and Euntai Kim},
  title = {{Interacting Multiple Model Vehicle-Tracking System Based on Neural Network}},
  journal = {Journal of Korean Institute of Intelligent Systems},
  year = {2009},
  doi = {10.5391/jkiis.2009.19.5.641},
}
Figure