Deep Learning Methods for Speed Estimation of Bipedal Motion from Wearable IMU Sensors.

Area of Science:

• Robotics
• Human-Computer Interaction
• Machine Learning

Background:

• Estimating human motion speed from wearable inertial measurement unit (IMU) sensors is crucial for applications like pedestrian dead reckoning.
• Traditional methods often rely on handcrafted features, which may not capture the complexity of human movement.

Purpose of the Study:

• To evaluate deep learning methods for predicting human motion speed using raw IMU data.
• To propose and validate a novel deep learning architecture for improved motion speed estimation.

Main Methods:

• Testing general-purpose deep learning architectures against classical feature-based approaches.
• Developing a novel semi-supervised variational auto-encoder with a specialized decoder (dense layer with sinusoidal activation).
• Experimenting with sensor placement on the shoe, shin, and thigh.

Main Results:

• Deep learning methods outperformed classical approaches in motion speed estimation.
• The proposed novel architecture achieved the lowest average error.
• Sensor placement on the shoe yielded the best results, with significant accuracy gains when using all three sensor locations.

Conclusions:

• Deep learning offers a powerful approach for estimating human motion speed from IMU data.
• The novel proposed architecture demonstrates superior performance for this specific task.
• Optimal sensor placement and multi-sensor fusion enhance estimation accuracy for pedestrian dead reckoning.