Novel robust computer vision algorithms for micro autonomous systems

Krerngkamjornkit, R 2014, Novel robust computer vision algorithms for micro autonomous systems, Doctor of Philosophy (PhD), Aerospace, Mechanical and Manufacturing Engineering, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Krerngkamjornkit.pdf Thesis application/pdf 3.40MB
Title Novel robust computer vision algorithms for micro autonomous systems
Author(s) Krerngkamjornkit, R
Year 2014
Abstract People detection and tracking are an essential component of many autonomous platforms, interactive systems and intelligent vehicles used in various search and rescues operations and similar humanitarian applications. Currently, researchers are focusing on the use of vision sensors such as cameras due to their advantages over other sensor types. Cameras are information rich, relatively inexpensive and easily available. Additionally, 3D information is obtained from stereo vision, or by triangulating over several frames in monocular configurations. Another method to obtain 3D data is by using RGB-D sensors (e.g. Kinect) that provide both image and depth data. This method is becoming more attractive over the past few years due to its affordable price and availability for researchers.

The aim of this research was to find robust multi-target detection and tracking algorithms for Micro Autonomous Systems (MAS) that incorporate the use of the RGB-D sensor. Contributions include the discovery of novel robust computer vision algorithms. It proposed a new framework for human body detection, from video file, to detect a single person adapted from Viola and Jones framework. The 2D Multi Targets Detection and Tracking (MTDT) algorithm applied the Gaussian Mixture Model (GMM) to reduce noise in the pre-processing stage. Blob analysis was used to detect targets, and Kalman filter was used to track targets. The 3D MTDT extends beyond 2D with the use of depth data from the RGB-D sensor in the pre-processing stage. Bayesian model was employed to provide multiple cues. It includes detection of the upper body, face, skin colour, motion and shape. Kalman filter proved for speed and robustness of the track management. Simultaneous Localisation and Mapping (SLAM) fusing with 3D information was investigated. The new framework introduced front end and back end processing.

The front end consists of localisation steps, post refinement and loop closing system. The back-end focus on the post-graph optimisation to eliminate errors.The proposed computer vision algorithms proved for better speed and robustness. The frameworks produced impressive results. New algorithms can be used to improve performances in real time applications including surveillance, vision navigation, environmental perception and vision-based control system on MAS.
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Aerospace, Mechanical and Manufacturing Engineering
Keyword(s) Artificial Intelligence
Autonomous System
Computer Vision
Detection and Tracking
RGB-D Sensor
Versions
Version Filter Type
Access Statistics: 254 Abstract Views, 453 File Downloads  -  Detailed Statistics
Created: Wed, 24 Jun 2015, 12:28:58 EST by Denise Paciocco
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us