RHflow4D - Optic Flow Computation using Robust Hough Technique with quasi
NAME
RHflow4D
RHflow4D
-r reference_frame
-c consecutive_frame
[
-b block_size
]
[
-o block_overlap
]
[
-f filter_size
]
[
-t confidence_threshold_level
]
[
-k kernel_type
]
[
-F output_filename
]
The RHflow4D is an implementation of the Robust Hough Transform Technique for Optic Flow estimation
and segmentation (Bober and Kittler 1993). It uses four parameter
motion model:
Vx = a1*x - a2*y + a3;
Vy = a2*x + a1*y + a4.
Motion parameters a3 and a4 describe the translation of the block center, whereas a1 and a2
correspond to "deformation" of the block (rotation, change of scale, etc).
Optic flow is estimated
with sub-pixel precision, up to 0.01 pixel for translation. The accuracy achieved in practice depends on the
displacements, type of texture, level of noise e.t.c. However, we have shown that the algorithm
is more accurate and robust then several state-of-the-art techniques. Several types of Robust Kernels
were implemented, including Absolute, Hampel_1_2_1, Andrews,
Tukey's. The optimal kernel
depends on the type of sequence and noise level and distribution. The algorithm uses multiresolution
optimization and is therefore fast, especially if non-overlapping blocks are used. The algorithm also
provides estimate accuracy measure associated with each pixel.
- -r reference_frame_filename
- Specifies the file name of the reference frame (stored in raster format).
- -c consecutive_frame_filename
- Specifies the file name of the consecutive frame (stored in raster format).
- -b block_size
- Defines the block size used for estimation. When the value b is specified, the actual block size
is 2*b+1
- -o block_overlap
- Specifies the block overlap: 1 - dense placement of the blocks, 2 - every second pixel, etc.
(if not specified, non-overlapping blocks assumed)
- -f binomial_filter_passes
- Defines the sigma of the low-pass Gaussian-like filter.
- -t confidence_threshold
- Defines the threshold value used for thresholding the static confidence measures. Option -F has to
be used to write the thresholded flow into a file.
- -k
- Specifies the kernel type:
0 - Absolute
1 - Hampel 1-2-1
2 - Andrew's sine kernel
3 - Tukey's kernel + quadratic error
4 - Experimental
- -F output_filename
- Write optic flow into file (Format Type 2). This format is used to evaluate estimate accuracy
(using program EstAccurcy).
- -h
- Print help
CHOOSING OPTIMAL PARAMETERS
RHflow4D -r treed.19 -c treed.20 -f 3 -b 7 -t 5000.0 -F flow.ff -
k 2 -o 1
Estimate optic flow using frames treet.19 (as reference) and treet.20 (as consecutive). Low-pass filter
both frames using binomial filter 3 passes. Use 15 times 15 rectangular blocks (2*7+1) and Andrew's
sine kernel. Threshold the flow at confidence level 5000.0 and write it to file <flow.ff>
For further information about the algorithms see:
M. Bober and J. Kittler, Robust motion analysis, CVPR1994, pp 947--952, IEEE Computer Society Press.
M. Bober and J. Kittler, Estimation of general multimodal motion: an approach based on
robust statistics and Hough transform, Image and Vision Computing, Volume 12, Number 12, 1994,
pp 661--668
RHflow4D, RHglob2D, RHglob4Dr, shs
During iterative procedure, the estimate of block displacement is not checked against valid block
positions. If displacement range is larger then the strip around image excluded from estimation,
segmentation error may result.
This program is copyrighted, see the Copy-Right Notice.
Miroslaw Bober, Vision Speech and Signal Processing Group,
University of Surrey, Guildford, Surrey, GU2 5XH, United Kingdom.
M.Bober@ee.surrey.ac.uk
03-Feb-98. Automatically converted by man2html,
written by G.Matas (g.matas@ee.surrey.ac.uk)