Lim, Siew Hooi (2008) Automatic markerless pose estimation for human in short sleeve or long sleeve attire. Masters thesis, Universiti Malaysia Sabah.
Markerless motion capture systems have the potential to provide an inexpensive, non-obtrusive solution for the estimation of body poses. Video analysis of human dynamics has a wide variety of applications in visual surveillance, visual-based human computer interaction, human robot interaction, gesture recognition and analysis, etc. Human pose estimation is a difficult problem, because human bodies are versatile, presenting a wide range of poses and self-occlusion occurrence. The goal of this project is to develop an automatic human pose estimation algorithm to recover the body and limb configurations in 20 from monocular video sequences without any special markers on the body. In this project, two approaches were described to model the human body poses from a monocular video sequence. In the first approach, a cardboard person model was proposed to model the upper part of human body using template matching algorithm. However silhouette feature alone was not enough to recover body configurations of the human when the body parts occluded each other and it was computationally expensive. Thus, a computer vision based approach was proposed as the second method to automatically detect human body parts and estimate the human body poses from a markerless monocular video sequence. The input image was first segmented using a silhouette extraction function based on the brightness level transformation to extract the moving silhouette patterns (human figures) from a static background for subsequent processing. Human body parts detection was then performed using colour, contours and silhouettes cues. Human body model initialization was performed in a fully automatic way. The only assumption was the person should be in an upright and frontal poses in the video sequence. Circular head fitting was first detected using circular Hough transform method. K-means clustering was then performed on the circular head region to obtain the skin colour distribution of the face. A skin colour model was built from the detected face region and used to find the candidate positions of limbs. The pixel classification performance was measured by using a receiver operating characteristic (ROC) curve. Radon transform was used to obtain more accurate orientation of the upper arms. Various physical and motion constraints regarding the human body were then used to construct the upper body configuration. The relation of the hand position inside the torso region was introduced to estimate the human pose for long sleeve attire's users. The final stage of this project was to recognize the pose performed by the human subject. Nineteen different body poses were considered for classification using feed-forward back propagation neural network, in which eight features are extracted from each pose. Our algorithm can recognize poses for any person entering the scene In either short sleeve or long sleeve shirt. It could also estimate the human poses even under illumination changes, self-occlusion occurrence and distance variations. There is no need for our system to use skin colour model built from skin pixel database for skin colour detection. Human pose estimation using computer vision-based approach reduces the computational cost by finding correct body part candidates more efficiently.
|Item Type:||Thesis (Masters)|
|Uncontrolled Keywords:||body pose, feed-forward backpropagation, configuration, motion capture system, algorithm|
|Subjects:||T Technology > TA Engineering (General). Civil engineering (General)|
|Divisions:||SCHOOL > School of Engineering and Information Technology|
|Deposited By:||IR Admin|
|Deposited On:||02 Dec 2014 09:30|
|Last Modified:||02 Dec 2014 09:30|
Repository Staff Only: item control page