Ranjbaran, Ali
(2013)
Improved Stereo Vision Algorithms For Robot Navigation.
PhD thesis, Universiti Sains Malaysia.
Abstract
The main motivation of this research is to find the best depth and direction for navigating a robot using stereo vision by solving the difficulties in finding disparity value for low information, noisy and tilted images as problem statement. An adaptive window method is implemented in three approaches. Approach 1 and 2 use common cost functions such as SSD, SAD and GB (gradient). Approach 3 uses a Linear-based function. By using adaptive method, the error in computing the disparity value for SSD is 12%, for Gradient-based is 10% and for Linear-based is 7%. SSD is 8 and 2 times faster than Gradient-based and Linear-based functions respectively. The linear-based technique with 50% more accurate than SSD is a suitable tool for stereo vision applications. The proposed denoising method for Lena and Barbara images are compared with ROF model by computing Total Variation (TV). When TV is 0.88 for noisy Lena image, ROF reduces TV to 0.21, Linear-based and Gradient-based decreases TV to 0.24 and 0.26 respectively. When TV is 0.88 for noisy Barbara image, TV is decreased to 0.68 by ROF, 0.62 by Linear-based and 0.65 by Gradient-based techniques. A stable platform system is designed and developed for stabilizing vision line horizontally for a moving robot. The system is equipped by a closed loop tilt and pan stabilizer using a dual-axis accelerometer sensor. It stabilizes the tilt and pan angles with 0.5 second time constant and steady state error lower that 0.025 radian. By using adaptive stereo matching that uses linear cost function, the quality of the best direction for navigation is improved to 88% of success rate.
Actions (login required)
 |
View Item |
