On the Functional and Extra-Functional Properties of IMU Fusion Algorithms for Body-Worn Smart Sensors.

Nils Büscher,Christian Haubelt,Daniel Gis,Volker Kühn

doi:10.3390/s21082747

Abstract

In this work, four sensor fusion algorithms for inertial measurement unit data to determine the orientation of a device are assessed regarding their usability in a hardware restricted environment such as body-worn sensor nodes. The assessment is done for both the functional and the extra-functional properties in the context of human operated devices. The four algorithms are implemented in three data formats: 32-bit floating-point, 32-bit fixed-point and 16-bit fixed-point and compared regarding code size, computational effort, and fusion quality. Code size and computational effort are evaluated on an ARM Cortex M0+. For the assessment of the functional properties, the sensor fusion output is compared to a camera generated reference and analyzed in an extensive statistical analysis to determine how data format, algorithm, and human interaction influence the quality of the sensor fusion. Our experiments show that using fixed-point arithmetic can significantly decrease the computational complexity while still maintaining a high fusion quality and all four algorithms are applicable for applications with human interaction.

Highlights

Inertial sensors using the Micro-Electro-Mechanical Systems (MEMS) technology have become the de-facto standard for inertial measurement units (IMU) in consumer electronics [1]
Our experiments show that using fixed-point arithmetic for the sensor fusion can drastically reduce the computational effort while still rendering the algorithms usable in most cases
A key aspect of the sensor fusion algorithms that are examined in the work at hand are their extra-functional properties

Summary

Introduction

Inertial sensors using the Micro-Electro-Mechanical Systems (MEMS) technology have become the de-facto standard for inertial measurement units (IMU) in consumer electronics [1] Due to their capabilities and energy efficiency, they are predestined for gesture- and activity recognition, health monitoring [2], smart clothes, or remote devices powered through energy harvesting. Before the evaluation of the functional and extra-functional properties of the sensor fusion algorithms are described in Sections 4 and 5, this section will provide general information about the used sensor fusion algorithms, data formats, hardware, and the implementation. This information is viable to put the results and interpretations in the correct context. The implementations of the extended Kalman filter as well as the complementary filter use the method described in [18] to determine the orientation from accelerometer data and magnetometer data to avoid the usage of computationally expensive trigonometric functions.

Objectives

Results

Discussion

Conclusion