Accelerometry-based inference of constrained motions
FRQNT Grant 2019-2023
( Ed. note: See the poster from the AI and physics conference of 2019. )
Accelerometer sensors, originally developed in the 1960’s to provide inertial-navigation capabilities for aircraft, ships and spacecraft, have now, with the advent of micro-electromechanical (MEMS) technology, become ubiquitous low-cost components providing motion-sensing capability in a wide variety of areas. In the field of biomedical applications alone, a partial list includes behavioural biometrics[1]; gesture recognition[2]; surgical skills[3]; detection of gait disorders[4] [5]; iOT-enabled systems for rehabilitation[6], assisted living[7] and elderly care[8]; and physiological monitoring[9].
The challenge for researchers lies no longer in the collection of the data but in its interpretation. One fundamental difficulty is the separation of the gravitational contribution to the accelerometer signal. This is often alleviated by the addition of a gyroscope to the system. Typically, the gyroscope is used to detect changes in orientation and the accelerometer is relegated to the detection of changes in linear velocity. The analysis is still vulnerable to another widely-recognized problem, viz. the drift error that accumulates when the orientation or position of an object is reconstructed through single or double-integration techniques.
For completely arbitrary unconstrained motions, such as in the free motion of an aircraft, there is no means of overcoming these limitations. However, when there are known constraints imposed on the motion paths, they can be exploited in the reconstruction of the remaining degrees of freedom. Although this can reduce error and problem complexity, there does not appear to be a systematic study of such strategies. A simple starting point is a circular-motion investigation[10] that demonstrates, without the use of gyroscopes, the ability to recover both the radial and angular position of accelerometers with respect to the axis of rotation. This contrasts with gyroscope-free techniques for unconstrained motions which requires anywhere from 12 to 18 accelerometers for a single body[11] [12].
In the proposed project, we will identify and explore simple low-degree-of-freedom systems that are sufficiently constrained so as to allow the motion to be fully and reliably inferred using only a single 3-axis accelerometer. The development of such techniques also promises broad applicability as a low-level classification layer in machine-learning (ML) architectures for motion in complex systems such as the human body[3][4]. Recognizing that traditional ML systems suffer from inflexibility and poor transfer to related domains, there is increasing interest in hybrid systems that incorporate domain-specific knowledge[15][16]. Automated detection or ruling-out of simple low-level constraints in a motion signal can serve as a universal feature extractor[17] converting a complex time series into a low-dimensional feature vector, into a single classification or some type of interval-dependent qualifier[18][19]. Different deep networks specialized for distinct application domains can re-use the same early stage, thus decreasing development time, increasing learning transfer opportunities and comprehensibility of the internal representations.
References
[1] A. Alzubaidi and J. K. Kalita, “Authentication of Smartphone Users Using Behavioral Biometrics,” IEEE Commun. Surv. Tutor., vol. 18, pp. 1998–2026, 2016.
[2] F. Hong, S. You, M. Wei, Y. Zhang, and Z. Guo, “MGRA: Motion Gesture Recognition via Accelerometer,” Sensors, vol. 16, no. 4, Apr. 2016.
[3] A. Krause, “Kalman filtering for real-time orientation tracking of handheld microsurgical instrument,” The Robotics Institute Carnegie Mellon University. .
[4] D. Slijepcevic et al., “Automatic Classification of Functional Gait Disorders,” IEEE J. Biomed. Health Inform., vol. 22, no. 5, pp. 1653–1661, Sep. 2018.
[5] I. H. López-Nava and A. Muñoz-Meléndez, “Towards Ubiquitous Acquisition and Processing of Gait Parameters,” in Advances in Artificial Intelligence, G. Sidorov, A. H. Aguirre, and C. A. R. García, Eds. Springer Berlin Heidelberg, 2010, pp. 410–421.
[6] I. Bisio, A. Delfino, F. Lavagetto, and A. Sciarrone, “Enabling IoT for In-Home Rehabilitation: Accelerometer Signals Classification Methods for Activity and Movement Recognition,” IEEE Internet Things J., vol. 4, pp. 135–146, 2017.
[7] T. R. Bennett, J. Wu, N. Kehtarnavaz, and R. Jafari, “Inertial Measurement Unit-Based Wearable Computers for Assisted Living Applications: A signal processing perspective,” IEEE Signal Process. Mag., vol. 33, pp. 28–35, 2016.
[8] D. Figo, P. C. Diniz, D. R. Ferreira, and J. M. P. Cardoso, “Preprocessing techniques for context recognition from accelerometer data,” Pers. Ubiquitous Comput., vol. 14, pp. 645–662, 2010.
[9] R. Khusainov, D. Azzi, I. E. Achumba, and S. D. Bersch, “Real-Time Human Ambulation, Activity, and Physiological Monitoring: Taxonomy of Issues, Techniques, Applications, Challenges and Limitations,” in Sensors, 2013.
[10] C. I. Larnder and B. Larade, “On the determination of accelerometer positions within host devices,” Am. J. Phys., p. ( in 2nd review ), 2018.
[11] E. Edwan, S. Knedlik, and O. Loffeld, “Constrained Angular Motion Estimation in a Gyro-Free IMU,” IEEE Trans. Aerosp. Electron. Syst., vol. 47, no. 1, pp. 596–610, Jan. 2011.
[12] Y. M. Al-Rawashdeh and M. Elshafei, “Filtering Techniques for Estimating the Angular Motion Using All-Accelerometers,” Applied Mechanics and Materials, 2016. [Online]. Available: https://www.scientific.net/AMM.829.103. [Accessed: 14-Oct-2018].
[13] L. Lo Presti and M. La Cascia, “3D Skeleton-based Human Action Classification: a Survey,” Pattern Recognit., vol. 53, Dec. 2015.
[14] P. Wang, W. Li, P. Ogunbona, J. Wan, and S. Escalera, “RGB-D-based Human Motion Recognition with Deep Learning: A Survey,” ArXiv171108362 Cs, Oct. 2017.
[15] K. Bousmalis, G. Trigeorgis, N. Silberman, D. Krishnan, and D. Erhan, “Domain Separation Networks,” ArXiv160806019 Cs, Aug. 2016.
[16] O. M. H. Rindal, T. M. Seeberg, J. Tjønnås, P. Haugnes, and Ø. Sandbakk, “Automatic Classification of Sub-Techniques in Classical Cross-Country Skiing Using a Machine Learning Algorithm on Micro-Sensor Data,” Sensors, vol. 18, no. 1, Dec. 2017.
[17] S.-A. Rebuffi, H. Bilen, and A. Vedaldi, “Efficient parametrization of multi-domain deep neural networks,” ArXiv180310082 Cs Stat, Mar. 2018.
[18] U. Mori, A. Mendiburu, E. Keogh, and J. A. Lozano, “Reliable Early Classification of Time Series Based on Discriminating the Classes over Time,” Data Min Knowl Discov, vol. 31, no. 1, pp. 233–263, Jan. 2017.
[19] R. E. Schapire, “The boosting approach to machine learning: An overview,” in Nonlinear estimation and classification, Springer, 2003, pp. 149–171.
[20] D. Baraff, “Linear-Time Dynamics Using Lagrange Multipliers,” in SIGGRAPH, 1996.
[21] K. Sherif, K. Nachbagauer, and W. Steiner, “On the rotational equations of motion in rigid body dynamics when using Euler parameters,” Nonlinear Dyn., vol. 81, no. 1–2, pp. 343–352, 2015.
[22] C. I. Larnder, “Laboratoires de physique partagés par le biais d’impression 3D,” Entente Canada-Québec relative à l’enseignement dans la langue de la minorité et à l’enseignement des langues secondes., 2018.