1. INTRODUCTION Thanks to our collaboration with AB-Horizon, we are developing a smart mirrored AR cobot to assist people during training. Our aim is to develop an intelligent vision system for action recognition in fitness training. Human movements during physical exercise are various and heterogeneous so traditional mechanical measurement systems are ill-suited to assess human motion in such a wide variety of movements [1]. A subjective approach is limited and not suitable for an extensive commercial production. However new techniques as Deep Learning (DL) for RGB images let us perform model-based estimation to measure quantities intrinsically subjective [2]. DL visual network reads data from the image plane and generates a skeletal model in this 2D representation. New DL architectures and frameworks are constantly released, their oriented approach to hardware optimization supports and inspires embedded device application [3][4]. In this paper we show our first result, the operative strategy and the calibration test we intend to perform to validate our pose estimation project. This project is focused on developing an embedded device capable of qualitative bio-mechanical analysis. A key point of this application is to establish a trade-off between hardware cost and software computational consumption: for this reason we benchmarked different solutions that could be used for this task. Different MPU suitable for inference engine (CPU, GPU, VPU) were tested with well-known DNN (Deep neural network) to assess their performance. Our specific objective leads to a highly customized and specialized software development: general on the motion to acquire and specific on the hardware constrains. We will confront various network manager such the TensorFlow framework and different optimization technique as Transfer Learning and Fine tuning for our sequential convolutional network. After choosing the right hardware, customized training operation for our model will be mandatory to accomplish accuracy requirements.

DEEP LEARNING FOR GESTURE RECOGNITION IN SPORTIVE TRAINING OPERATION PERFORMED BY STANDALONE SPECIALIZED NEURAL NETWORK VISION-BASED SYSTEMS

Bernardo Lanza
;
Cristina Nuzzi;Simone Pasinetti;Matteo Lancini
2021

Abstract

1. INTRODUCTION Thanks to our collaboration with AB-Horizon, we are developing a smart mirrored AR cobot to assist people during training. Our aim is to develop an intelligent vision system for action recognition in fitness training. Human movements during physical exercise are various and heterogeneous so traditional mechanical measurement systems are ill-suited to assess human motion in such a wide variety of movements [1]. A subjective approach is limited and not suitable for an extensive commercial production. However new techniques as Deep Learning (DL) for RGB images let us perform model-based estimation to measure quantities intrinsically subjective [2]. DL visual network reads data from the image plane and generates a skeletal model in this 2D representation. New DL architectures and frameworks are constantly released, their oriented approach to hardware optimization supports and inspires embedded device application [3][4]. In this paper we show our first result, the operative strategy and the calibration test we intend to perform to validate our pose estimation project. This project is focused on developing an embedded device capable of qualitative bio-mechanical analysis. A key point of this application is to establish a trade-off between hardware cost and software computational consumption: for this reason we benchmarked different solutions that could be used for this task. Different MPU suitable for inference engine (CPU, GPU, VPU) were tested with well-known DNN (Deep neural network) to assess their performance. Our specific objective leads to a highly customized and specialized software development: general on the motion to acquire and specific on the hardware constrains. We will confront various network manager such the TensorFlow framework and different optimization technique as Transfer Learning and Fine tuning for our sequential convolutional network. After choosing the right hardware, customized training operation for our model will be mandatory to accomplish accuracy requirements.
File in questo prodotto:
File Dimensione Formato  
Bernardo_Lanza - estesa.doc

accesso aperto

Licenza: Dominio pubblico
Dimensione 1.4 MB
Formato Microsoft Word
1.4 MB Microsoft Word Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11379/558875
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ateneo

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact