Human pose estimation is the process that aims to locate body parts and build human body representations from input data such as images and video. It is typically a computationally difficult operation, where, in order to achieve accurate results, the use of expensive GPUs is mandatory. Nowadays new use cases, such as augmented reality, demand to make this kind of operations viable on mobile and edge devices and research in other fields, such as human-robot collaboration, is leaning towards building portable and inexpensive solutions. This thesis describes the design and prototyping process of a real-time human pose estimation network made using edge devices, building a network using only Raspberry Pi boards for image processing, exploiting the TensorFlow lite library for running the necessary Deep Convolutional Neural Network components and utilizing the Robot Operating System framework to build a fast, real-time system. Additionally some techniques for creating DCNN that are capable of real-time execution will also be discussed and evaluated in order to try and surpass the hardware limitations imposed by the setup. A solution that archives real-time results with good accuracy is obtained with this work.
Human pose estimation is the process that aims to locate body parts and build human body representations from input data such as images and video. It is typically a computationally difficult operation, where, in order to achieve accurate results, the use of expensive GPUs is mandatory. Nowadays new use cases, such as augmented reality, demand to make this kind of operations viable on mobile and edge devices and research in other fields, such as human-robot collaboration, is leaning towards building portable and inexpensive solutions. This thesis describes the design and prototyping process of a real-time human pose estimation network made using edge devices, building a network using only Raspberry Pi boards for image processing, exploiting the TensorFlow lite library for running the necessary Deep Convolutional Neural Network components and utilizing the Robot Operating System framework to build a fast, real-time system. Additionally some techniques for creating DCNN that are capable of real-time execution will also be discussed and evaluated in order to try and surpass the hardware limitations imposed by the setup. A solution that archives real-time results with good accuracy is obtained with this work.
Real-time multi-camera 3D human pose estimation on edge devices
SAVOIA, EMANUELE FRANCESCO
2023/2024
Abstract
Human pose estimation is the process that aims to locate body parts and build human body representations from input data such as images and video. It is typically a computationally difficult operation, where, in order to achieve accurate results, the use of expensive GPUs is mandatory. Nowadays new use cases, such as augmented reality, demand to make this kind of operations viable on mobile and edge devices and research in other fields, such as human-robot collaboration, is leaning towards building portable and inexpensive solutions. This thesis describes the design and prototyping process of a real-time human pose estimation network made using edge devices, building a network using only Raspberry Pi boards for image processing, exploiting the TensorFlow lite library for running the necessary Deep Convolutional Neural Network components and utilizing the Robot Operating System framework to build a fast, real-time system. Additionally some techniques for creating DCNN that are capable of real-time execution will also be discussed and evaluated in order to try and surpass the hardware limitations imposed by the setup. A solution that archives real-time results with good accuracy is obtained with this work.File | Dimensione | Formato | |
---|---|---|---|
Emanuele_Francesco_Savoia.pdf
accesso aperto
Dimensione
6.33 MB
Formato
Adobe PDF
|
6.33 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/80172