Pedestrian Navigation Method Based on Machine Learning and Gait Feature Assistance

Abstract

In recent years, as the mechanical structure of humanoid robots increasingly resembles the human form, research on pedestrian navigation technology has become of great significance for the development of humanoid robot navigation systems. To solve the problem that the wearable inertial navigation system based on micro-inertial measurement units (MIMUs) installed on feet cannot effectively realize its positioning function when the body movement is too drastic to be measured correctly by commercial grade inertial sensors, a pedestrian navigation method based on construction of a virtual inertial measurement unit (VIMU) and gait feature assistance is proposed. The inertial data from different positions of pedestrians' lower limbs are collected synchronously via actual IMUs as training samples. The nonlinear mapping relationship between inertial information from the human foot and leg is established by a visual geometry group-long short term memory (VGG-LSTM) neural network model, based on which the foot VIMU and virtual inertial navigation system (VINS) are constructed. The VINS experimental results show that, combined with zero-velocity update (ZUPT), the integrated method of error modification proposed in this paper can effectively reduce the accumulation of positioning errors in situations where the gait type exceeds the measurement range of the inertial sensors. The positioning performance of the proposed method is more accurate and stable in complex gait types than that merely using ZUPT.

Keywords: gait feature assistance; gait phase recognition; machine learning; pedestrian navigation; virtual inertial navigation system.

Figure 1

The overall scheme of pedestrian navigation method (a) Construction of neural network; (b) Diagram of the pedestrian navigation process.

Figure 2

Side view and front view of human lower rigid-body kinematics model.

Figure 3

VGG-LSTM model architecture.

Figure 4

Distributed structure for inertial information acquisition.

Figure 5

Gyroscope and accelerometer information from pedestrian foot (a) Gyroscope information from pedestrian foot; (b) Accelerometer information from pedestrian foot.

Figure 6

Gyroscope and accelerometer information from pedestrian thigh (a) Gyroscope information from pedestrian thigh; (b) Accelerometer information from pedestrian thigh

Figure 7

Information comparison of virtual accelerometers with reference accelerometers (a) Comparison of virtual accelerometer X axis information with reference accelerometer X axis information; (b) Comparison of virtual accelerometer Y axis information with reference accelerometer Y axis information; (c) Comparison of virtual accelerometer Z axis information with reference accelerometer Z axis information.

Figure 8

Information comparison of virtual gyroscope with reference gyroscope (a) Comparison of virtual gyroscope X axis information with reference gyroscope X axis information; (b) Comparison of virtual gyroscope Y axis information with reference gyroscope Y axis information; (c) Comparison of virtual gyroscope Z axis information with reference gyroscope Z axis information.

Figure 9

Normal gait cycle of pedestrian.

Figure 10

Virtual and actual IMU data under the walking velocity of 1 and 2.5 m/s: (a) X axis virtual and actual gyroscope outputs with a walking velocity of 1 m/s; (b) Y axis virtual and actual accelerometer outputs with a walking velocity of 1 m/s; (c) X axis virtual and actual gyroscope outputs with a walking velocity of 2.5 m/s; (d) Y axis virtual and actual accelerometer outputs with a walking velocity of 2.5 m/s.

Figure 11

Performance verification experiments of pedestrian navigation methods: (a) Indoor pedestrian navigation experiment route; (b) Comparison of indoor curves by different pedestrian navigation methods; (c) Outdoor pedestrian navigation experiment route; (d) Comparison of outdoor curves by different pedestrian navigation methods.