"Speed": A Dataset for Human Speed Estimation.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Inertial Measurement Units (IMUs) are increasingly used for human motion analysis.
• Accurate speed estimation is crucial for various applications, including sports science and rehabilitation.
• Existing speed estimation techniques using IMUs have limitations in accuracy and generalizability.

Purpose of the Study:

• To introduce a new, medium-scale dataset named the "Speed" dataset for IMU-based speed estimation.
• To benchmark existing speed estimation methods on this dataset.
• To develop and evaluate a deep learning model, SpeedNet, for improved subject-independent and subject-specific speed estimation.

Main Methods:

• Collected walking/running speed data (4.0-9.5 km/h) from 33 healthy subjects using IMUs.
• Created a dataset with accelerometer and gyroscope data, organized into subject-independent sets.
• Implemented and benchmarked Support Vector Regression, Gaussian Process Regression, and shallow neural networks.
• Designed and tested a deep Convolutional Neural Network (CNN) named SpeedNet.

Main Results:

• SpeedNet achieved a subject-independent Root Mean Square Error (RMSE) of 0.4819 km/h.
• Subject-specific adaptation of SpeedNet reduced the RMSE to 0.1747 km/h.
• SpeedNet demonstrated superior accuracy and real-time performance compared to other benchmarked methods.

Conclusions:

• The "Speed" dataset provides a valuable resource for IMU-based speed estimation research.
• SpeedNet offers a highly accurate and computationally efficient solution for estimating human walking/running speeds.
• The proposed subject-specific adaptation significantly enhances the precision of speed estimation.