Abstract

Owing to the possibility of skidding with the ground in motion, the positioning of the autonomously-navigated automated guided vehicle (AGV) based on the wheel odometry (Odom) will bring errors, meanwhile, the rotation error is remarkably larger than the linear motion error. Three neural network methods were proposed in this study for the data fusion of the Odom and IMU, i.e., back-propagation (BP) neural network method, one-dimensional convolutional (1D-CNN) method, and long short-term memory (LSTM) method. The data fusion results of the Odom and IMU obtained by the extended Kalman filtering method were taken as a reference. The performance of the aforementioned three neural networks for the data fusion of the Odom and IMU was compared respectively. The experimental results indicate that all three neural networks can reduce the rotation error of AGV to some extent. It was also found that 1D-CNN possesses the fastest training speed, and the positioning accuracy is the highest when employing 1D-CNN for data fusion in the application.

中文翻译：

---

基于神经网络的室内自主导航AGV运动轨迹数据融合

摘要

由于运动中可能与地面打滑，基于车轮里程计（Odom）的自主导航自动导引车（AGV）的定位会带来误差，同时旋转误差明显大于直线运动误差. 本研究针对Odom和IMU的数据融合提出了三种神经网络方法，即反向传播（BP）神经网络方法、一维卷积（1D-CNN）方法和长短期记忆（LSTM）方法。 ） 方法。以扩展卡尔曼滤波方法得到的Odom和IMU的数据融合结果作为参考。分别比较了上述三种神经网络在 Odom 和 IMU 数据融合方面的性能。实验结果表明，三种神经网络都能在一定程度上降低AGV的旋转误差。还发现1D-CNN训练速度最快，应用中采用1D-CNN进行数据融合时定位精度最高。