Pattern Recognition Methods as a Tool to Build an Automatic System for Learning Coordinated Human Motions

Abstract

There is great social and economic significance in the teaching and learning of motion activities. This is notably true for teaching the activities involved in rehabilitation, sports, and professional work. The possibility of engaging an automatic teaching system is highly significant. Nevertheless, building an effective system is an ongoing challenge. This article describes a general outline of the teaching system, which includes MEMS (micro-electro-mechanical systems) sensors, haptic actuators, and algorithms for signal classification applied to the online selection of an appropriate teaching method. The main goal of this paper was to prove that the system is able to teach fast and synchronized movements effectively. To this end, system performance was presented and discussed. The statistical tests revealed an efficiency of the proposed approach, especially for tasks of teaching fast and periodic movements. This result was the primary outcome of the presented paper. The described scheme can be utilized for building two types of motor learning systems. The first relates to the "personal" learning systems for rehabilitation and sports. The second type can perform the classification of complex movements of human body parts and may be used in teaching the remote control of machines and vehicles (excavators, cranes, search and rescue drones, etc.).INDEX TERMS haptic feedback, machine learning, MEMS sensors, motor learning, pattern recognition.

Machine Learning Methods Used in the Automatic System for Teaching Human Motions—Key Aspects of CNN, HMM, and Minimum Distance Algorithms

“…In order to select and activate the appropriate teaching method in a given situation, the classification of motion signals must be performed [10], [13]. For example, a specific method (algorithm) can be used to learn precise movements of the hand holding a surgical instrument, but in certain situations (identified by the classification process), it can be replaced by an algorithm that improves coordination of both hands.…”

“…As examples, the following classification methods can be mentioned here: Support Vector Machine (SVM) [14], [15], methods based on Artificial Neural Networks (ANNs) in particular: Convolutional Neural Network (CNN), Convolutional Recurrent Neural Network (CRNN) [2], [14], [16]- [20], methods using Hidden Markov Model (HMM) [21], minimum distance methods, e.g., k-Nearest Neighbor (kNN) [16], [22], syntactic pattern recognition methods [23], [24], and other methods using special and mixed solutions [25]- [29]. The above methods are finding increasing use in areas of technology and the humanities that have not previously been associated with the automatic interpretation of signals [13], [30]- [33].…”

“…The classification methods were tested using a set of exemplary motion signals recorded during the basic tests of the learning system [10]. They consisted in examining the general idea of the automatic selection of a teaching algorithm depending on the results of the motion signals classification [10], [13]. In this experiment, the participants performed a movement relating to a circular arms motion during swimming butterfly stroke -Fig.…”

“…The second class C 1 corresponds to major motion errors, which require a more decisive reaction of the teaching system. This is done by sending a vibration signal at the calculated moment in the motion period [7], [10], [13]. The signal informs the learner about the need to interrupt an incorrect sequence of motion.…”

“…This average result can be calculated in various ways. For example voting techniques can be implemented (see details in [13]). The consequence of this process is that the requirements for the quality of classification are not very high.…”