Learning Action Primitives From 3D Stereo Vision Measurements

Arango Slingsby, Ricardo

4. term

Kandidatuddannelse, Computer Vision and Graphics (Ballerup)

2010

I computer vision er det vigtigt at forstå, hvordan objekter bevæger sig i billeder. Mange eksisterende metoder beskriver hver bevægelsesbane for sig, selv når forskellige baner deler fælles delstræk. Det skaber redundant og ukorrelerede data. Vi kalder disse delte delstræk for handlingsprimitiver—grundlæggende byggeklodser for bevægelse—og der er brug for modeller, der kan forbinde dem. Denne afhandling præsenterer en ramme, der visuelt sporer objekter ved hjælp af farvesegmentering (gruppering af pixels efter farve) og Hidden Markov Models (sandsynlighedsmodeller for sekvenser). Rammen registrerer bevægelsesbaner, identificerer handlingsprimitiver og bygger en samlet, kompakt model, der fælles repræsenterer flere baner via deres delte primitiver. Modellen over handlingsprimitiver kan bruges til robotlæring og giver en højere-niveau beskrivelse af udførte handlinger. Rammen er implementeret som et C++‑programmeringsbibliotek med en komplet løsning til objektdetektion, sporing, bevægelsesregistrering og modelopbygning.

In computer vision, understanding how objects move in images is important. Many approaches describe each motion trajectory separately, even when different trajectories share common segments. This leads to redundant, uncorrelated data. We refer to these shared segments as action primitives—basic building blocks of motion—and we need models that can link them. This thesis presents a framework that visually tracks objects using color segmentation (grouping pixels by color) and Hidden Markov Models (probabilistic models for sequences). It records motion trajectories, identifies action primitives, and builds a single, compact model that jointly represents multiple trajectories through their shared primitives. The action-primitive model can be used for robot learning, providing a higher-level description of the actions performed. The framework is implemented as a C++ programming library with a complete system for object detection, tracking, motion recording, and model construction.

[This abstract was generated with the help of AI]

Action Primitives ; Color Segmentation ; Blob Filtering ; Hidden Markov Models ; Imitation Learning ; Stereo Tracking ; Motion Trajectories

Download
View record in AAU Student Projects