Five-Direction Occlusion Filling with Five Layer Parallel Two-Stage Pipeline for Stereo Matching with Sub-Pixel...

Area of Science:

• Computer Vision
• Robotics
• Embedded Systems

Background:

• Binocular stereoscopic matching is crucial for depth perception in computer vision.
• Semi-Global Matching (SGM) is a popular algorithm known for efficiency and accuracy.
• Existing SGM algorithms struggle with accuracy in long-range applications.

Purpose of the Study:

• To propose a novel disparity improvement strategy for SGM.
• To enhance accuracy and robustness in long-range stereo matching.
• To develop a hardware-efficient architecture for real-time applications.

Main Methods:

• Subpixel interpolation and disparity optimization post-processing.
• Area optimization, hardware-friendly divider, and split look-up table implementation.
• Clock alignment multi-directional disparity occlusion filling and floating-point depth acquisition.

Main Results:

• Achieved a non-occlusion error rate of 4.61% on the KITTI2015 dataset, outperforming state-of-the-art.
• Developed a hardware architecture on Stratix-IV using 5.6 K LUTs, 12.8 K registers, and 2.5 M bits memory.
• Reached a maximum frequency of 98.28 MHz for 640x480 resolution, processing at 320 FPS with 1.459 W power consumption.

Conclusions:

• The proposed strategy significantly improves SGM accuracy, especially for long-range depth estimation.
• The optimized hardware architecture offers a compelling balance of performance, power, and resource utilization.
• This work advances real-time, high-accuracy stereo vision applications.