Stereo Matching with Double Refinements for Centralized Texture Depth Packing and Their VLSI Implementations

  • 周 敬倫

Student thesis: Master's Thesis

Abstract

In this thesis a stereo matching with double refinements which combines centralized texture depth packing (CTDP) system is proposed In the traditional stereo matching algorithm left-right check in the refinement process is unable to detect mismatching errors usefully In order to raise the accuracy texture-based detection is proposed in this thesis to flag mismatching errors for achieving high accuracy In addition the CTDP format which has no corresponding hardware implementation is proposed recently This thesis proposed the CTDP system combined with a stereo matching with double refinements The proposed stereo matching algorithm changes the reference pixels in the census transform Then computes initial matching costs by Hamming distances of bit-streams resulted with new selected reference pixels In the cost aggregation the proposed stereo matching algorithm a hardware-friendly adaptive support weights are adopted to perform iterative aggregations for generating better disparities Finally double refinements are used to repair the error disparities in boundary and occluded regions For the CTDP format both texture and depth are down-scaled by Lanczos filter for packing with CTDP format The corresponding VLSI architecture of stereo matching and CTDP format are provided and realized in Altera FPGA The design of stereo matching requires 27k logic elements 30k registers and 4 8Mb RAM and the process speed achieves 60 frames per second with the resolution of 1920 × 1080 and 64 disparity levels in 160 MHz The design of CTDP architecture requires 5 776 logic elements 7 791 registers and 1 43 Mb RAM and the process speed achieves 60 frames per second with the resolution of 1920 × 1080
Date of Award2016 Aug 18
Original languageEnglish
SupervisorBin-Da Liu (Supervisor)

Cite this

'