Development of saliency metric for autonomous landmark selection in cognitive robot navigation / Gao Hanyang

Gao , Hanyang (2022) Development of saliency metric for autonomous landmark selection in cognitive robot navigation / Gao Hanyang. Masters thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (210Kb)
	PDF (Thesis M.A) Download (2032Kb)

Abstract

Urban landmarks are spatial features that are visually significant in the neighbourhood. Humans cognitively select landmarks based on their visual appearance like size, colour, and shape. Many researchers have attempted to evaluate and extract visual landmarks, usually by abstracting their features and quantifying their salience. Humans use qualitative, high level visual features of landmarks in their navigation. In contrast, robots use empirical, low-level HOG and SURF features in their landmark extraction for navigation. A quantitative model for visual salience indicators in urban landmark extraction seems beneficial to the robotics community and could improve understanding for cognitive robot navigation. Quantifying visual salience indicators for urban landmark extraction is challenging when the goal is to compute qualitative, high-level visual features. Existing robot landmark extraction methods are based on low-level features like HOG and SURF, which fails to express landmarks cognitively like humans. This dissertation proposes an algorithm to quantify urban landmarks based on visual salience indicators for cognitive robot navigation. The dissertation follows three objectives; to segment urban landmarks in an image, to develop an algorithm to quantify visual salience indicators for urban landmarks extraction, and to compare the performance of proposed algorithm in extracting urban landmarks between robot and human. A drone is used to collect fourteen aerial images of urban landmarks. Four images are taken from top view and for variation, one image is taken from front view, for each landmark. The images processing follows bilateral filtering, Otsu thresholding, morphing to resolve connectedness issues, and segmenting the landmarks. Next, the size, colour and shape salience equations are considered following pixel counting, extracting intensity value from hue, saturation and value (HSV), and an equation for shape indicator, respectively. The experiment done suggests that the final salience value for each landmark can be calculated by adding size, colour and shape together according to weightage 45%, 35% and 20% respectively. Sixty participants between the age of 18 and 60 agree to answer a survey in evaluating 14 urban landmarks based on their size, colour and shape. Encouragingly, 12 out of the 14 urban landmarks selected by the robot match the human selection, with 85.7% accuracy.

Actions (For repository staff only : Login required)