Evaluation of Skeleton Trackers and Gesture Recognition for Human-robot Interaction.
Translated title
Evaluering af menneske- og gestikulerings-genkendelse til menneske-robot interaktion.
Author
Bünger, Martin
Term
4. term
Publication year
2013
Submitted on
2013-06-06
Pages
83
Abstract
Dette kandidatspeciale inden for vision, grafik og interaktive systemer dokumenterer to semestres arbejde ved Georgia Institute of Technology. Rapporten har to dele. Først evalueres skeletsporing (software der estimerer en persons kropsstilling ud fra et Kinect-dybdekamera) ved at sammenligne Microsofts og PrimeSense's sporingssystemer. Testene viser, at deres ydeevne er meget ens. Der afprøves også en simpel opsætning med flere Kinect-sensorer og datafusion (sammenkobling af data fra flere enheder), som giver små forbedringer. Dernæst undersøges gestusgenkendelse med skjulte Markov-modeller, en statistisk metode til at analysere tidsafhængige sekvenser, med fokus på at genkende pegegester. Forsøgene omfatter både to-klasse- og multiklasse-opsætninger, og begge giver gode resultater.
This master's thesis in vision, graphics, and interactive systems documents two semesters of work at the Georgia Institute of Technology. The report has two parts. First, it evaluates skeleton trackers (software that estimates a person's body pose from a Kinect depth camera) by comparing the Microsoft and PrimeSense tracking systems. Tests show that their performance is very similar. The study also tries a simple setup with multiple Kinect sensors and data fusion (combining information from more than one device), which gives slight improvements. Second, it investigates gesture recognition using hidden Markov models, a statistical approach for analyzing sequences over time, with a focus on detecting pointing gestures. Experiments include both two-class and multi-class setups, and both perform well.
[This abstract was generated with the help of AI]
Keywords
Kinect ; Microsoft ; OpenNI ; Primesense ; NITE ; gestikulering
Documents