Hierarchical extreme learning machine for body part detector

• Main Author: Lim, Zhong Hui
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/67451
• Institution: Nanyang Technological University

In the recent years, body part detection had gain significant interest and popular because of the capability of its impact towards understanding human behavior under distinctive circumstances. Estimating and marking of human body parts of different posture in videos or pictures give fundamental insights for breaking down human conduct and activities. Nevertheless, Human body exhibits huge amount of variation in an image due to variation in posture, viewing angle etc making body part detection very difficult. Currently, body posture estimation are limited as it is only applicable to human with upright posture. With the aim to detect non-upright postures, it is helpful to distinguish separate body parts. Taking into account the perceived areas of head and torso, we can deduce the body action either upright or non-upright position. The first part of the project start by pre-processing images that include collecting image dataset using Google web crawler after which the images will be appropriately crop and resize to use as training images. Each training image will consist of a unique features which would be extracted via Histogram of Oriented Gradient (HOG) method. Subsequently, the features that was extracted will then be used to train the HELM classifier, creating a model for each body part. This model will be utilized as database of H-ELM classifier for classifying/detection of testing images. The testing images are from video dataset collected and being extracted into frames and afterward extract the foreground using bounding boxes. The features on the foreground will then be segmented using mean shift segmentation and feed into the H-ELM classifier for classification. By using the position of head and torso that had been detected as a reference, the posture of the person inside the frame can then be estimated as upright posture, non-upright posture or not a human. Another project which uses H-ELM classifier to detect whole person as well as estimating the posture was used to compare with the posture estimation using whole person detector and results had proven that posture estimation is more suitable to use in the framework of body part detector.