Improved Position Estimation Algorithm of Agricultural Mobile Robots Based on Multisensor Fusion and Autoencoder Neural Network

Peng Gao, Hyeonseung Lee, Chan-Woo Jeon, Changho Yun, Hak-Jin Kim, Weixing Wang, Gaotian Liang, Yufeng Chen, Zhao Zhang, Xiongzhe Han2022 · DOI

TL;DR

This paper introduces a novel multisensor fusion algorithm for agricultural mobile robots (AMRs) that leverages an autoencoder-Radial Basis Function (ARBF) neural network to dynamically optimize the noise parameters of an Extended Kalman Filter (EKF). This approach significantly improves position estimation accuracy, especially in challenging environments where traditional GNSS signals are unreliable, and demonstrates robustness even when sensors partially fail.

What Problem Does This Paper Address?

Agricultural mobile robots (AMRs) are increasingly vital for precision agriculture tasks like planting, weeding, and harvesting. Accurate real-time position estimation is critical for their autonomous operation and to ensure they follow precise paths without damaging crops or missing targets. While Global Navigation Satellite Systems (GNSS), particularly RTK-GNSS, offer high precision, their performance degrades significantly in environments with signal interference from buildings or dense foliage. Inertial Measurement Units (IMUs) and odometry sensors can operate independently but accumulate errors over time. Therefore, a robust multisensor fusion approach is needed to overcome the limitations of individual sensors and provide reliable positioning, even under adverse conditions.

Methods

The core of this work is an improved position estimation algorithm that combines multisensor data from RTK-GNSS, IMU, and dual rotary encoders using an Extended Kalman Filter (EKF). The novelty lies in the optimization of the EKF's state and measurement noise matrices (Q and R). This optimization is achieved through a proposed Autoencoder-Radial Basis Function (ARBF) neural network.

Here's a breakdown:

  • Multisensor Fusion: Data from RTK-GNSS (high-precision positioning), IMU (attitude and motion data), and rotary encoders (wheel speed) are fused.
  • Extended Kalman Filter (EKF): A standard EKF is used to combine the sensor data and predict the AMR's state (position, orientation, velocity, etc.).
  • ARBF Neural Network for Noise Optimization:
    • Autoencoder: This unsupervised learning component is used for feature extraction and denoising of the raw sensor data. It learns to represent the data efficiently and reconstruct it, effectively reducing noise.
  • Radial Basis Function (RBF) Network: The denoised data is then fed into an RBF network, which models the noise sequence. This allows for the dynamic estimation of the noise matrices (Q and R) for the EKF.
  • Two-Dimensional Motion Model: The AMR is simplified to a 2D motion model for state estimation.

Datasets

The paper does not explicitly mention pre-existing datasets for training or evaluation. Instead, the authors constructed an AMR test platform and conducted experiments in various real-world environments (road, grass, field) at Kangwon National University. They collected data from their integrated sensors during these experiments.

Key Insights and Ideas

  • Dynamic Noise Optimization is Crucial: The paper strongly emphasizes that fixed noise matrices in EKF can limit performance, especially in dynamic and noisy environments. The ARBF approach dynamically adjusts these parameters, leading to better adaptation.
  • Autoencoders for Robust Denoising: Using autoencoders for denoising sensor data before feeding it into the fusion algorithm is a powerful technique for improving signal quality and reducing the impact of environmental noise.
  • ARBF's Effectiveness in Adverse Conditions: The ARBF algorithm demonstrates significant improvements not only in ideal conditions but also when RTK-GNSS signals are interfered with (simulated using metal plates) or when sensors like encoders fail. This highlights its robustness.
  • IMU's Indispensable Role: Experiments with sensor failures clearly show that while encoders provide velocity, the IMU is critical for obtaining heading angle and acceleration data, which are essential for accurate position prediction in the fusion system.

Key Results

  • Improved Accuracy: The proposed ARBF algorithm achieved position estimation accuracies that were up to 45.9% better than RTK-GNSS alone in dynamic experiments.
  • Static Performance: In static experiments, the ARBF algorithm reduced CEP and 2DRMS values by up to 25.9% and 24.4%, respectively, compared to RTK-GNSS.
  • Robustness to Interference: Even when RTK-GNSS signals experienced interference, the ARBF algorithm's RMSE was significantly lower (0.245 m) compared to the interfered RTK-GNSS (1.216 m) and even better than single-GNSS (0.482 m).
  • Sensor Failure Tolerance: The system showed improved performance even when encoders failed. However, it was noted that IMU failure led to significant position deviations, underscoring its importance.
  • Denoising Impact: Table 1 shows that the ARBF algorithm effectively reduced the standard deviation (Std) of various sensor data dimensions, indicating noise suppression, while having minimal impact on the mean values.

Limitations and Future Work

  • Computational Resources: The authors acknowledge that the ARBF algorithm requires significant computing resources, suggesting future work could involve deploying the model on a cloud server for remote interactions to improve real-time performance.
  • Heading Angle Prediction: The current system does not explicitly predict the heading angle, which is noted as important for navigation. Future work could focus on incorporating heading angle prediction into the multisensor fusion algorithm.
  • Data Output Frequency: To meet the demands of high-speed navigation, further research could explore increasing the overall data output frequency, potentially by improving edge computing capabilities.
  • Vision Navigation Integration: Combining the proposed system with vision navigation is suggested as a future direction, especially given the increasing application of vision-based navigation.
  • Interference Scenarios: The paper suggests studying multisensor fusion algorithms for situations with prolonged RTK-GNSS or GNSS signal interruptions, such as in tunnels.
  • Static Experiment Site Variation: The authors mention that static experiments were conducted at sites with different horizontal planes, leading to slight variations in results. Future studies could use identical environmental parameters for better comparison.

Summary updated

BibTeX
@misc{gao2022improved,
  author = {Peng Gao and Hyeonseung Lee and Chan-Woo Jeon and Changho Yun and Hak-Jin Kim and Weixing Wang and Gaotian Liang and Yufeng Chen and Zhao Zhang and Xiongzhe Han},
  title = {{Improved Position Estimation Algorithm of Agricultural Mobile Robots Based on Multisensor Fusion and Autoencoder Neural Network}},
  year = {2022},
  doi = {10.3390/s22041522},
}
Figure