3D Human Pose Estimation from Monocular Image Sequences

Barbulescu, Adela

4. term

Vision, Graphics and Interactive Systems, Master

2012

At rekonstruere 3D-positurer for mennesker ud fra billedsekvenser fra et enkelt kamera er svært, men nyttigt i mange sammenhænge. Mange eksisterende metoder bruger maskinlæring som Support Vector Machines (SVM) eller Gaussiske processer, men de kæmper ofte i rodede scener og kan kræve ekstra input som silhuetter eller strengt kontrollerede kameraopsætninger. Dette projekt præsenterer en ramme, der estimerer en persons 3D-positur uden baggrundsinformation fra enkeltkamera-billedsekvenser og er robust over for variationer i kameraindstillinger. Rammen modellerer den iboende ikke-linearitet i menneskelig bevægelse med fleksible læringskomponenter: en stærkt tilpasselig 2D-detektor til at finde kropsspor i hvert billede og en Gaussisk proces-regressor trænet på specifikke handlinger, som løfter disse 2D-spor til 3D. På HumanEva benchmark overgik systemet tidligere arbejder med en 70 % reduktion i den gennemsnitlige estimeringsfejl på de samme datasæt. Vi giver detaljer om forsøgsopsætninger, testresultater og præstationsmål for 3D-positursestimering.

Reconstructing 3D human body poses from images taken with a single camera is difficult but useful in many areas. Many existing methods use machine learning techniques such as Support Vector Machines or Gaussian processes, yet they often struggle in cluttered scenes and may need extra inputs like silhouettes or tightly controlled camera setups. This project presents a framework that estimates a person’s 3D pose from sequences of single-camera images without background information and that remains robust when camera settings vary. It models the non-linear nature of human motion with flexible learning components: a highly customizable 2D detector to find body cues in each frame, and a Gaussian process regressor trained on specific action motions to map these 2D cues into 3D. On the HumanEva benchmark, the system outperformed previous work, achieving a 70% reduction in average estimation error on the same datasets. We provide the experimental settings, test results, and performance measures for 3D pose estimation.

[This abstract was generated with the help of AI]

3D human pose estimation, support vector machine, part-based model, gaussian process

Download
View record in AAU Student Projects

Barbulescu, Adela:

• 3D Human Pose Estimation from Monocular Image Sequences (2012)