Zaznavanje gest v video tokovih na vgrajeni napravi

Rolih, Blaž (author), Čehovin Zajc, Luka (mentor)

V okviru diplomskega dela je implementiran in ovrednoten nosljiv prototip za zaznavanje ročnih gest, ki deluje na vgrajeni napravi OAK-D iz platforme DepthAI. Vgrajena naprava omogoča učinkovit zajem slike in obdelavo le-te z uporabo raznih operacij računalniškega vida, vključno z izvajanjem globokih nevronskih mrež. Prototip z uporabo več zaporednih nevronskih mrež in vmesnih operacij določi pozicijo roke v prvoosebnem načinu, roki sledi in glede na časovni potek pozicije roke določi gesto. Vse to skoraj v celoti teče na vgrajeni napravi, kar razbremeni gostiteljski sistem in omogoča nizke zakasnitve pri zaznavanju. Za praktično testiranje je implementirano upravljanje predvajalnika glasbe. S tem namenom je zbrana podatkovna množica gest, ki kljub svojemu omejenemu obsegu omogoča, da se sistem zanesljivo nauči prepoznavati različne geste. Sistem je eksperimentalno evalviran na testni množici, kjer dosega zaželeno točnost. Dobro se obnese tudi v realnem scenariju, kjer je bil sistem preizkušen s strani testnih uporabnikov z upravljanjem glasbe v realnem času.

• Language: Slovenian
• Keywords: geste, računalniški vid na vgrajenih napravah, DepthAI, CNN, LSTM
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2022
• PID: 20.500.12556/RUL-139723
• COBISS.SI-ID: 121799939
• Publication date in RUL: 06.09.2022

ROLIH, Blaž, 2022, Zaznavanje gest v video tokovih na vgrajeni napravi [online]. Bachelor’s thesis. [Accessed 23 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=139723

Secondary language

• Language: English
• Title: Gesture recognition in video streams on an embedded device

Abstract

In this diploma thesis, a wearable prototype for the detection of hand gestures is implemented and evaluated, which works on the OAK-D embedded device from the DepthAI platform. Embedded device is capable of efficient image capture and image processing using various computer vision operations, including deep neural networks. Using a sequence of neural networks and intermediate operations, the prototype determines the position of the hand in first-person mode, tracks the hand and, based on the time course of the hand position, determines the gesture. All of this runs almost entirely on the embedded device, offloading the host system and enabling low detection latencies. For practical testing, music player control is implemented. For this purpose, a dataset of gestures has been collected, which, despite its limited scope, enables the system to reliably learn to recognize different gestures. The system is experimentally evaluated on a test set, where it achieves the desired accuracy. It also performs well in a real-world scenario where the system has been tested by test users controlling music playback in real-time.

• Keywords: gestures, embedded computer vision, DepthAI, CNN, LSTM