**5. Conclusion**

*Sports Science and Human Health - Different Approaches*

**4.2 AI for motor-control quantification**

data (big data), and theoretical understanding.

and analyses [44, 52, 53, 67].

for researchers and practitioners.

reliable massive training data.

**of 3D training big data**

sufficient to apply the class mechanics to quantitatively determine the model system. Currently, AI could alleviate this challenge due to its "learning" ability [44, 64–66].

Since the inputs of the limited IMUs cannot mechanically determine the twochain model, AI technologies are the alternative ways for the two-chain model quantification. Studies have shown that AI techniques have become a powerful tool for helping to solve many challenging problems in human motor-skill evaluations

The basic idea of AI prediction is to find a way to learn general features of existing data in order to make sense of new data [64, 65]. This description highlights the central role of data for establishing implicit knowledge. The amount of data must be sufficiently large to provide many training examples from which a large set of parameters can be extracted. In the past decades, AI techniques have experienced a resurgence following concurrent advances in computer power, large amounts of

Among the AI technologies, deep learning is considered as a powerful tool that percolates through to all application areas of AI, such as image identification, speech recognition, natural language processing, and, indeed, biofeedback support [68–70]. The success of deep learning networks encourages their implementation in further applications for the enhancement of human physical activities [52, 54, 67]. Recently, *Nature Neuroscience* has published the latest developments in the area of markerless, video-based motion tracking, indicating that the power of deep learning will enable motion tracking to human-like accuracy [53]. This study confirms that motion capture/quantification of limbs' coordination will move from an expensive and difficult task restricted to the laboratory to an effortless daily routine

From motor learning point of view, wearables would have much higher potential than video shooting in the future practice. This is not only because of the fast advance in miniature of wearables but also due to three inherited drawbacks of video-shooting approach, i.e., (1) the limited capture space, (2) the complexity of capture systems (from setup, calibration, to operation), and (3) the time-consuming nature of data processing (high cost of data processing). Reliable biomechanical feedback should be obtained from accurate quantification of human movement in field, with some requiring large space. Even with a multi-camera setting, unexpected environmental factors (e.g., interactions among athletes) will create a data gap. Further, it is true that we are already sitting on massive movement data (e.g., YouTube, Flickr) for training of deep learning models; but the video datasets are uncalibrated and have very little information on the hardware and conditions used to capture particular videos, which can bias the deep learning recognition algorithms [71]. Currently, the availability of reliable motion capture data for developing deep learning models is significantly limited. In summary, the combination of the two-chain full-body model with six wearable IMUs and the deep learning prediction based on IMUs' data shows great potential in developing real-time biomechanical feedback training for an efficient human motor-skill learning and optimization. The missing piece for testing the potential is

**4.3 The key for raising the reliability of wearables: creating a diversification** 

Two factors revealed by previous studies strongly influence deep learning performance [65, 66, 72–74]. One is the massive data, and the other one is the

**48**

This chapter highlights the challenges and future of wearable technology in human motor-skill learning and optimization. It introduces a novel two-chain biomechanical body model with six IMUs that are powered with deep learning technology. The framework can serve as a basis for developing real-time biomechanical feedback training in practice. In order to create a universal biomechanical feedback device for learning and training of any human motor skill, the massive and diverse big data of multifarious human motor skills have to be created first. One realistic way for obtaining the big data is through a synchronized measurement from 3D motion capture and IMUs. Evidently, gaining high-quality, full-body motion data across sports and arts performances would currently be the vital step for the realtime biomechanical feedback development.

The realization of the methodological breakthrough will allow us to transform the human motor learning paradigm from a largely subjective art into a precise scientific method. The potentials would be to (1) take scientific monitoring of motor skills from a lab-based environment into the field; (2) simplify a scientific movement quantification, transitioning from using a complicated motion capture system to easily applied wearables; and (3) transfer the vital biomechanical feedback in real time to prevent the worst/movement errors from happening while finding individual compensation/optimization. This methodology is the culmination of research programs in biomechanics, anthropometry, computer sciences, pedagogy, and equipment development. It aims to build innovative technologies for generating new knowledge as well as practical and definitive scientific methods for

empowering motor learning. Fulfillment and application of the new wearable-based method in the future will benefit diverse human physical activities, including, but not limited to, motor-skill acquisition in sports, arts performances, health/fitness, and recreational activities.
