A Neural architecture for recognising human actions in video sequences

Malan, Christian (2022-04)

Thesis (MEng)--Stellenbosch University, 2022.

Thesis

ENGLISH ABSTRACT: The goal of an action recognition model (as applied to a video) is to extract distinct features that accurately characterise the action present and then from that determine the dominant action in the video. Since such actions occur in the spatio-temporal domain, recognising it relies on both spatial and temporal features. In this thesis we designed and built such an action recognition model by making use of a deep neural network system that combines various subsystems, each focussing on complemen tary aspects of the problem. A particular focus is on the automatic extraction of discriminatory spatio-temporal features. In combination the final system is shown to accurately distinguish between a limited range of actions.

AFRIKAANSE OPSOMMING: Die doel van ’n aksieherkenningsmodel (soos toegepas op ’n video) is om eerstens duidelike kenmerke te onttrek wat die aksie akkuraat beskryf, en dan daaruit die dominante aksie teenwoordig in die video te bepaal. Aangesien sulke aksies in beide tyd en ruimte plaasvind, moet ’n herkenner beide van hierdie domeine in ag neem. In hierdie tesis het ons ’n diep neuralenetwerk stelsel vir hierdie doel ontwikkel. Dit maak gebruik van verskeie substelsels wat elk fokus op komplementêre aspekte van die probleem. ’n Besondere fokus is op die outomatiese onttrekking van diskriminerende tydruimtelike kenmerke. In kombinasie is die finale stelsel in staat om ’n beperkte reeks aksies akkuraat te onderskei.

Please refer to this item in SUNScholar by using the following persistent URL: http://hdl.handle.net/10019.1/124921
This item appears in the following collections: