5 The Probabilistic Occupancy Map is a procedure which estimates the
6 marginal probabilities of presence of individuals at every location
7 in an area of interest under a simple appearance model, given binary
8 images corresponding to the result of a background-subtraction from
11 The appearance model is parametrized by a family of rectangles which
12 approximate the silhouettes of individuals standing at every
13 location of interest, from every point of view.
18 This program should compile by typing `make' on any standard
19 GNU/Linux system, as long as the libpng is installed with its header
22 To run a simple test on the images provided in the archive just type
26 It will generate several result files in the /tmp directory.
28 If the program is not provided with a filename, it will process the
29 standard input, but will not make any output.
34 The configuration file contains the following keywords
36 ROOM <view width> <view height> <number of cameras> <number of locations>
38 Defines the input image size, the number of cameras and the number
39 of locations in the area of interest.
41 RECTANGLE <camera number> <location number> notvisible|<xmin> <ymin> <xmax> <ymax>
43 Defines the parameters of a certain rectangle, standing for an
44 individual at a certain location viewed from a certain camera. All
45 non-specified rectangles are "not visible" by default.
47 INPUT_VIEW_FORMAT <input image filename format>
49 Specifies the filenames of the input images produced by the
50 background subtraction. The immobile parts should be black (0, 0,
51 0) and the moving ones white (255, 255, 255). See FILENAME FORMAT
54 RESULT_VIEW_FORMAT <result image filename format>
56 Specifies the filenames of the result images. See FILENAME FORMAT
59 RESULT_FORMAT <result filename format>
61 Specifies the filenames of the result probability files, which
62 contains one marginal probability per line, hence as many lines as
63 there are locations of interest. See FILENAME FORMAT below.
65 CONVERGENCE_VIEW_FORMAT <convergence file name format>
67 Specifies the filenames of the images for individual iteration
68 during the convergence of the algorithm. Use with care, since tens
69 of images will be produced for every single frame. See FILENAME
74 Sets the prior probability of presence (common to all
75 locations). Default is 0.01.
79 Sets the variance of the distance between the image produced by
80 the background subtraction and the ideal synthetic image. Default
83 MAX_NB_SOLVER_ITERATIONS
85 Sets a bound on the number of iterations. Default is 100.
89 Sets the probability of absence ignored by the solver to speed up
90 the convergence (use with care, it can produce false
91 positive). Default is 1.0
93 The file test.pom provided in the archive gives an example. Lines
94 starting with "#" are ignored.
99 Every filename format in the configuration file is a string with the
100 following three "conversion specifications":
108 Some of these fields may be meaningless, depending with the
109 context. For instance the iteration number is defined only for
110 CONVERGENCE_VIEW_FORMAT.
115 For more information about the POM algorithm, please check the
116 section V and appendix A of:
118 François Fleuret, Jérôme Berclaz, Richard Lengagne and Pascal Fua,
119 "Multi-Camera People Tracking with a Probabilistic Occupancy Map",
120 IEEE Transactions on Pattern Analysis and Machine Intelligence
121 (TPAMI), 2007, to be published.
123 This distributed version of the software is slightly slower than the
124 one described in the article, for which many values such as the
125 image size were hard-coded in the source to let the compiler
131 This source code is available under the terms of the version 3 of
132 the GNU General Public License as published by the Free Software
133 Foundation. In short: If you distribute a software that uses POM,
134 you have to distribute it under GPL version 3, hence with the source
135 code. Another option is to contact us to purchase a commercial
141 Please mail pom@epfl.ch for bug reports, comments and questions.