Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
A generic computer vision human activity recognition tool.
Thesis (MEng)--Stellenbosch University, 2023.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2023
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614016884441088 |
|---|---|
| access_status_str | Open Access |
| author | Jacobs, Johannes Jacobus |
| author2 | Nel, Gerrit Stephanus |
| author_browse | Jacobs, Johannes Jacobus Nel, Gerrit Stephanus |
| author_facet | Nel, Gerrit Stephanus Jacobs, Johannes Jacobus |
| author_sort | Jacobs, Johannes Jacobus |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch : Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2023. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/127046 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:45:20.375Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2023 |
| publishDateRange | 2023 |
| publishDateSort | 2023 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/127046 A generic computer vision human activity recognition tool. Jacobs, Johannes Jacobus Nel, Gerrit Stephanus Stellenbosch University. Faculty of Engineering. Dept. of Industrial Engineering. Computer vision Human activity recognition Decision support systems Thesis (MEng)--Stellenbosch University, 2023. ENGLISH ABSTRACT: It is well-known that computers possess the capabilities to outperform humans in a variety of tasks, ranging from basic tasks, such as arithmetic calculations, to more advanced tasks, such as solving complex optimisation problems. Humans, however, are able to employ their biological senses (e.g. touch, smell, and vision) to perform tasks that are significantly more challenging for computers to perform — until recently, that is. Contemporary advances in the domain of artificial intelligence and the increased capabilities of computer hardware have led to the computational viability and proliferation of computer vision — the field of study that algorithmically enables computers to “see” and comprehend a physical environment. A prominent application area within the domain of computer vision is human activity recognition. Various powerful approaches towards human activity recognition employ computer vision algorithms to automatically detect people’s actions from video footage. Examples of human activity recognition tasks include video activity classification, human pose estimation, action detection, and action localisation, to name a few. Most computer vision algorithms are deep learning based approaches, which refer to multiple-layer artificial neural networks. The deep learning architectures investigated in this thesis towards performing human activity recognition are convolutional neural networks and residual neural networks. In this thesis, a generic decision support tool capable of performing human activity recognition by employing computer vision is developed. The objective of the decision support tool is to facilitate the analysis of video footage by learning to identify human activities from video footage. The decision support tool facilitates the processing of raw data, the training of a computer vision model in respect of the processed data, and the deployment of the trained computer vision model in respect of unseen video footage. An investigation into the pertinent literature related to deep learning and computer vision, as well as the literature pertaining to human activity recognition and the fundamental concepts of decision support tools in general, are presented. A computerised implementation of the decision support tool is designed, developed, and applied to a benchmark data set in order to verify the functionally correct working of the tool. As part of the computerised implementation, a graphical user interface is designed and developed. Furthermore, the decision support tool is also implemented with respect to a case study data set provided by an industry partner, which relates to video footage of a cash replenishment process at automated teller machines. Parameter tuning is performed by means of a sensitivity analysis to determine the best parameter value combination for the human activity recognition model with respect to the case study data set. The decision support tool acts as a proof of concept which the industry partner can utilise to monitor the quality of their cash replenishment process (in terms of process breaches) in an automated manner. During the case study, the surveillance footage is analysed by identifying and visualising the actions performed in the footage, yielding high-quality results. The results are also validated by various subject matter experts. AFRIKAANS OPSOMMING: Dit is welbekend dat rekenaars oor die vermo¨e beskik om beter as mense te presteer in ’n verskeidenheid take, wat wissel van basiese take, soos rekenkunde-tipe berekeninge, tot meer gevorderde take, soos die oplos van komplekse optimeringsprobleme. Mense is egter in staat om hul biologiese sintuie (bv. aanraking, reuk, en visie) te gebruik om take uit te voer wat aansienlik meer uitdagend is vir rekenaars om uit te voer — tot onlangs. Hedendaagse vooruitgang in die vakgebied van kunsmatige intelligensie en die verhoogde vermo¨ens van rekenaarhardeware het gelei tot die berekeningslewensvatbaarheid van rekenaarvisie — die studieveld wat rekenaars algoritmies in staat stel om ’n fisiese omgewing te “sien” en te verstaan. ’n Prominente toepassingsgebied binne die vakgebied van rekenaarvisie is menslike aktiwiteitsherkenning. Verskeie kragtige benaderings tot die herkenning van menslike aktiwiteit maak gebruik van rekenaarvisie-algoritmes om mense se optrede outomaties op te spoor vanaf videomateriaal. Voorbeelde van menslike aktiwiteitherkenningstake sluit in video-aktiwiteitklassifikasie, menslike postuurberaming, ksiebespeuring, en aksielokalisering, om ’n paar te noem. Die meeste rekenaarvisie-algoritmes is diepleergebaseerde benaderings, wat verwys na kunsmatige neurale netwerke met veeltallige lae. Die diepleer-argitekture wat in hierdie tesis ondersoek word vir die uitvoering van menslike aktiwiteitsherkenning, is konvolusionele neurale netwerke en residuele neurale netwerke. In hierdie tesis word ’n generiese besluitsteunhulpmiddel ontwikkel wat in staat is om menslike aktiwiteitsherkenning uit te voer deur middel van rekenaarvisie. Die doel van die besluitsteunhulpmiddel is om die ontleding van videomateriaal te fasiliteer deur te leer om menslike aktiwiteite uit videomateriaal te identifiseer. Die besluitsteunhulpmiddel fasiliteer die verwerking van rou data, die opleiding van ’n rekenaarvisiemodel ten opsigte van die verwerkte data, en die ontplooiing van die opgeleide rekenaarvisiemodel ten opsigte van ongesiene videomateriaal. ’n Ondersoek na die toepaslike literatuur, wat verband hou met diepleer en rekenaarvisie, asook die literatuur met betrekking tot menslike aktiwiteitsherkenning en die fundamentele konsepte van besluitsteunhulpmiddel in die algemeen, word aangebied. ’n Gerekenariseerde implementering van die besluitsteunhulpmiddel word op ’n maatstafdatastel toegepas om die funksionele korrekte werking van die instrument te verifieer. As deel van die gerekenariseerde implementering word ’n grafiese gebruikerskoppelvlak ontwerp en ontwikkel. Verder word die besluitsteunhulpmiddel ook ge¨ımplementeer met betrekking tot ’n gevallestudie-datastel wat deur ’n bedryfsvennoot verskaf word en verband hou met videomateriaal van die kontantaanvullingsproses by outomatiese tellermasjiene. Parameterinstelling word uitgevoer deur middel van ’n sensitiwiteitsanalise om die beste parameterwaardekombinasie vir die menslike aktiwiteitherkenningsmodel met betrekking tot die gevallestudiedatastel te bepaal. Die besluitsteuninstrument dien as ’n bewys van konsep, wat die bedryfsvennoot kan gebruik om die kwaliteit van hul kontantaanvullingsproses (in terme van prosesoortredings) op ’n outomatiese wyse te monitor. Tydens die gevallestudie word die toesigbeeldmateriaal ontleed deur die aksies wat in die beeldmateriaal uitgevoer word te identifiseer en te visualiseer, wat ho¨e kwaliteit resultate lewer. Die resultate word ook deur verskeie vakkundiges bekragtig. Masters 2023-02-10T17:01:57Z 2023-05-18T07:01:40Z 2023-02-10T17:01:57Z 2023-05-18T07:01:40Z 2023-02 Thesis http://hdl.handle.net/10019.1/127046 en_ZA en_ZA Stellenbosch : Stellenbosch University xxiv, 140 pages : illustrations. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Computer vision Human activity recognition Decision support systems Jacobs, Johannes Jacobus A generic computer vision human activity recognition tool. |
| title | A generic computer vision human activity recognition tool. |
| title_full | A generic computer vision human activity recognition tool. |
| title_fullStr | A generic computer vision human activity recognition tool. |
| title_full_unstemmed | A generic computer vision human activity recognition tool. |
| title_short | A generic computer vision human activity recognition tool. |
| title_sort | generic computer vision human activity recognition tool |
| topic | Computer vision Human activity recognition Decision support systems |
| url | http://hdl.handle.net/10019.1/127046 |
| work_keys_str_mv | AT jacobsjohannesjacobus agenericcomputervisionhumanactivityrecognitiontool AT jacobsjohannesjacobus genericcomputervisionhumanactivityrecognitiontool |