Implementation of visual saliency detection for autonomous rescue robot

• Main Author: Shantanu Singh.
• Other Authors: Wang Han
• Format: Final Year Project
• Language: English
• Published: 2010
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: http://hdl.handle.net/10356/41427
• Institution: Nanyang Technological University

As a part of development of sensory perception tools for an autonomous rescue robot, this project discusses and implements some of the saliency detection mechanisms that can be used. For this purpose, two important and established models of saliency detection were studied. The first model, the Itti-Koch-Niebur model, is a landmark algorithm that derives from the biologically inspired approach to saliency detection. This model was studied, implemented and tested in this project. Following this, two variations of this model were examined, one involving the use of LAB colour space, and the other using symmetry as a feature that contributes to salience. Both these extensions were implemented in OpenCV, and their results were analysed. The second major approach that was studied was the Spectral Residual approach, which is a purely computational method. This model was also implemented and its output saliency maps were studied. The saliency maps obtained from the four different approaches were finally compared, and the reasons for these differences were looked at. The results showed that the modified Itti-Koch-Niebur approach, with LAB colour space and Symmetry saliency was the clear winner in terms of accuracy. The LAB colour space was found to show significant improvement over RGB colour space. Further, although the symmetry maps added to the quality of the output saliency maps, they were also more time consuming and processing power intensive. Hence, the modified Itti-Koch-Niebur with LAB colour space and symmetry saliency proved to be the most effective approach to saliency detection, and was recommended for implementation on the robot.