Inimese keha pooside tuvastamine neurovõrkude abil kasutades klassipõhiste andmete augmenteerimist
Files
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Infotehnoloogia ja selle pidev arenemine on võimaldanud arvutitel näha ja õppida. Mida meie näeme oma silmadega, on võimalik jagada piksliteks ja anda pikslid sisendiks arvutile. Pikslite väärtuste põhjal saab arvuti näha ja õppida tundma ära erinevaid objekte, mida neile tundma õpetatakse. Arvutinägemisel ja õppimisel on palju võimalikke rakendusi. Antud lõputöös pakume välja raamistiku, mis suudab automaatselt tuvastada inimese keha poose traditsioonilise odava kaameraga tehtud pildilt. Meie lähenemisviis ühendab arvutinägemise ja neurovõrgud, et tuvastada inimene pildilt. Antud protsess algab silueti eraldamisega pildilt ning jätkub neurovõrgu kasutamisega. Neurovõrk tuvastab keha poosi eraldatud silueti põhjal. Suutmaks siluetti seostada keha poosiga, treeniti neurovõrku eelnevalt töödeldud piltidega siluettidest. Saadud tulemuste põhjal pakub antud raamistik paljulubavaid tulemusi aktsepteeritava täpsusega.
Information technologies and continuous development of information technologies have made it possible for computers to see and learn. What we see with our eyes, can be divided into pixels and fed to a computer, giving the computer the ability to see and learn based on the pixel values. Based on the input values computers can learn to recognize different objects depending on the examples taught to them. There are many possible applications for computers to see and learn in order to solve new tasks. In this thesis, we propose a framework, capable of automatically recognizing human body poses from a single image, obtained with a traditional low-cost camera. Our approach combines computer vision with neural networks to detect a human from an image. This process starts by extracting the silhouette from an image and then using a neural network to recognize body poses based on the extracted silhouettes. In order to match detected silhouettes with body poses, the neural network was trained with an already classified augmented dataset of preprocessed images depicting silhouettes. According to our results, we show that the proposed method provides promising results with acceptable accuracy.
Information technologies and continuous development of information technologies have made it possible for computers to see and learn. What we see with our eyes, can be divided into pixels and fed to a computer, giving the computer the ability to see and learn based on the pixel values. Based on the input values computers can learn to recognize different objects depending on the examples taught to them. There are many possible applications for computers to see and learn in order to solve new tasks. In this thesis, we propose a framework, capable of automatically recognizing human body poses from a single image, obtained with a traditional low-cost camera. Our approach combines computer vision with neural networks to detect a human from an image. This process starts by extracting the silhouette from an image and then using a neural network to recognize body poses based on the extracted silhouettes. In order to match detected silhouettes with body poses, the neural network was trained with an already classified augmented dataset of preprocessed images depicting silhouettes. According to our results, we show that the proposed method provides promising results with acceptable accuracy.