Verilog implementation of the census transform stereo vision algorithm.
This is based on work first published as "A parameterized stereo vision core for FPGAs" By Stephen Longfield and Mark Chang, FCCM09. Any academic work derived from this code should cite that paper.
As an example, with a maximum disparity of 100, and a window of 19x19 pixels, this generates the below depth map from the Middlebury Cones dataset:
Note that there are more errors around the edges of the image, where there's less to compare between the image pairs, as well as in areas where the image has fewer features.
To reproduce, run make venv
in the verilog
directory, followed by make gen_test
in
the verilog/stereo
directory.