Investigating the Path Tracking Algorithm Based on BP Neural Network

Lu Liu, Mengyuan Xue, Nan Guo, Zilong Wang, Yuwei Wang, Qixing TangSensors 2023 · DOI

TL;DR

This paper introduces a novel adaptive path tracking algorithm, BP-PP, that enhances the accuracy and robustness of autonomous vehicles, particularly in complex park environments. By integrating a Backpropagation (BP) neural network with the traditional Pure Pursuit (PP) algorithm, the system dynamically adjusts steering angles based on vehicle speed, path curvature, and lateral error, outperforming existing methods in both simulations and real-world tests.

What Problem Does This Paper Address?

Autonomous vehicles need to precisely follow planned paths, a critical task for safe and efficient operation. Traditional path tracking algorithms, like Pure Pursuit (PP), often struggle with varying road conditions, sudden changes in curvature, and the inherent non-linearities of vehicle dynamics. This can lead to significant tracking errors, especially in environments like industrial parks where obstacles, lane changes, and complex paths are common. The challenge is to develop an algorithm that is not only accurate but also robust to these dynamic and unpredictable factors, without requiring excessive computational resources.

Methods

The core of this work is the BP-PP algorithm, which enhances the classic Pure Pursuit (PP) algorithm. Here's how it works:

  • Kinematic Model & PP Foundation: The paper starts with a standard two-wheel kinematic model of a vehicle. The PP algorithm is used to derive an initial front-wheel steering angle based on a look-ahead point on the reference path.
  • BP Neural Network Integration: A Backpropagation (BP) neural network is introduced to refine the PP algorithm's output. The network is trained using data collected from simulations and real-world driving, with inputs including:
  • Vehicle speed
    • Radius of path curvature
    • Lateral error (distance from the vehicle to the planned path)
  • Adaptive Control Coefficient: The BP neural network outputs a control coefficient (Kc). This coefficient is then used to adjust the steering angle calculated by the PP algorithm, effectively creating an adaptive steering command. This allows the system to dynamically compensate for non-linear disturbances and changing road conditions.
  • Data Normalization: To improve the learning efficiency of the BP neural network, input data (vehicle speed, curvature radius, lateral error) is normalized to a range of [-1, 1].

Datasets

The paper utilizes a dataset generated from simulations and real-world experiments.

  • Simulated Data: Data was collected across various scenarios involving different vehicle speeds, path curvature radii, and lateral errors. This dataset was used to train the BP neural network.
  • Real Vehicle Data: Data was collected from an electric patrol vehicle equipped with GNSS/IMU sensors during experiments on rectangular and large curvature paths. This data was used for validation and to demonstrate the algorithm's performance in practical settings.

Key Insights and Ideas

  • Synergy of PP and BP: The combination of a geometric approach (PP) with a data-driven learning method (BP neural network) proves highly effective. PP provides a good baseline, while BP fine-tunes it to handle complex, non-linear dynamics.
  • Dynamic Adaptation is Key: The BP neural network's ability to learn and adapt based on real-time inputs like speed, curvature, and lateral error is crucial for improving tracking accuracy, especially in dynamic environments.
  • Robustness to Curvature Changes: The BP-PP algorithm shows a significant improvement in handling paths with changing curvatures, a common challenge for traditional algorithms that can lead to corner-cutting issues.
  • Reduced Interference of Non-linear Factors: The BP-PP approach effectively mitigates the impact of non-linear factors on path tracking, leading to more stable and accurate performance.

Key Results

The BP-PP algorithm demonstrated superior performance compared to traditional PP and Stanley algorithms in both simulations and real-world experiments:

  • Simulation Experiments:
    • Straight Path: Average tracking error reduced by 0.025 m at 3 m/s.
  • Curved Paths (1.5 m/s): Average tracking error reduced by 0.27 m (6.8 m radius), 0.42 m (5.5 m radius), and 0.67 m (4.5 m radius).
  • Real Vehicle Experiments:
    • Rectangular Road: Average tracking error reduced by 0.1 m.
  • Large Curvature Road: Average tracking error reduced by 0.086 m.
  • Response Speed: The BP-PP algorithm exhibited the fastest response speed in tracking initiation.
  • Robustness: The algorithm showed great robustness, particularly on high-curvature paths and in the presence of disturbances like deceleration bumps.

Limitations and Future Work

  • Limitations:
    • The paper notes that while vehicle kinematics methods offer some improvement, a limited increase in accuracy suggests that theory-based algorithms might be needed for further optimization.
  • The current model considers three main parameters affecting accuracy; however, in practice, more factors can influence path tracking.
  • Future Work:
    • Explore theory-based algorithms like Model Predictive Control (MPC) or Linear Quadratic Regulator (LQR) instead of kinematic models.
  • Investigate more complex working environments such as farmland and hilly terrain.
  • Integrate deep learning techniques into the system to further enhance path tracking accuracy and explore the factors influencing it.

Summary updated

Abstract

In this paper, we propose an adaptive path tracking algorithm based on the BP (back propagation) neural network to increase the performance of vehicle path tracking in different paths. Specifically, based on the kinematic model of the vehicle, the front wheel steering angle of the vehicle was derived with the PP (Pure Pursuit) algorithm, and related parameters affecting path tracking accuracy were analyzed. In the next step, BP neural networks were introduced and vehicle speed, radius of path curvature, and lateral error were used as inputs to train models. The output of the model was used as the control coefficient of the PP algorithm to improve the accuracy of the calculation of the front wheel steering angle, which is referred to as the BP-PP algorithm in this paper. As a final step, simulation experiments and real vehicle experiments are performed to verify the algorithm's performance. Simulation experiments show that compared with the traditional path tracking algorithm, the average tracking error of BP-PP algorithm is reduced by 0.025 m when traveling at a speed of 3 m/s on a straight path, and the average tracking error is reduced by 0.27 m, 0.42 m, and 0.67 m, respectively, at a speed of 1.5 m/s with a curvature radius of 6.8 m, 5.5 m, and 4.5 m, respectively. In the real vehicle experiment, an electric patrol vehicle with an autonomous tracking function was used as the experimental platform. The average tracking error was reduced by 0.1 m and 0.086 m on a rectangular road and a large curvature road, respectively. Experimental results show that the proposed algorithm performs well in both simulation and actual scenarios, improves the accuracy of path tracking, and enhances the robustness of the system. Moreover, facing paths with changes in road curvature, the BP-PP algorithm achieved significant improvement and demonstrated great robustness. In conclusion, the proposed BP-PP algorithm reduced the interference of nonlinear factors on the system and did not require complex calculations. Furthermore, the proposed algorithm has been applied to the autonomous driving patrol vehicle in the park and achieved good results.
BibTeX
@inproceedings{liu2023investigating,
  author = {Lu Liu and Mengyuan Xue and Nan Guo and Zilong Wang and Yuwei Wang and Qixing Tang},
  title = {{Investigating the Path Tracking Algorithm Based on BP Neural Network}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23094533},
}
Figure