Improving the Usability of Real-Time Passive Stereo Sensors
Muhammad Mehran Arshad Khan · ResearchSpace (University of Auckland) · 2014
There is an increasing industrial demand for robust real-time intelligent vision systems. High resolution stereo can play an important role but the high computational complexity for generating a depth map is the main concern. Hardware acceleration (using GPU or FPGA) can help satisfy real-time requirements. Symmetric Dynamic Programming Stereo (SDPS) is the highest throughput algorithm available on hardware (both in GPU and FPGA). However, the depth maps are generally noisy which reduces its usability. Therefore, the target here is to design and propose fast post-processing algorithms for the data generated by SDPS hardware which will improve its usability and will assist high level vision tasks. This thesis makes three main contributions. The first one is the fast depth contour generation algorithm. Depth contours represent object shapes which is the key feature for object modelling, detection, pose estimation and recognition. The second contribution is the Points-of-Interest detection algorithm; it is fast and robust to depth noise due to its use of left and right image pair. Both contour generation and Points-of-Interest detection take one pass which reduces the search space and then uses the reduced search space for subsequent processing. The third contribution are the two depth contour refinement algorithms. To improve depth contours, conventionally depth maps are refined before generating contours which can remove some fine details present in depth map. Two novel contour refinement algorithms are presented in this thesis which improve depth contours and reduce the number of streaks in the contour while preserving fine detail. Keywords: Real-time stereo, symmetric dynamic programming stereo (SDPS), depth contours, points of interest detection, key points detection, depth features, depth contour refinement, streak points count, contour mapping measure.