owlps/TODO.t2t

342 řádky
9.6 KiB
Plaintext
Surový Trvalý odkaz Blame Historie

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

Work to do in OwlPS
= Documentation =
- Known Doxygen problems:
- Struct member fields are not documented with INLINE_SIMPLE_STRUCTS
http://bugs.debian.org/725122
- Ugly NAME (new line) in libraries' man pages
http://bugs.debian.org/630020
- Missing spaces around const
http://bugs.debian.org/630018
- Documentation pages to be added:
- owlps-basics.t2t: basics of indoor positioning (definitions:
reference point, etc.), common command-line options, etc.
- owlps-experimenting.t2t: useful information gathered from the
previous experiments.
http://git.lm7.fr/?p=owlps-experiments.git
- Possibly owlps-autocalibration.t2t to expand the section
“Autocalibration algorithms” of owlps-positionerd.t2t. Sources of
information: PhD thesis, IPIN 2013 article.
= Global =
- Known bugs:
- libconfuse bug (Listener & Aggregator): http://bugs.debian.org/639115
- Check the size of the received packets to avoid buffer overflows.
- Use select(2) or poll(2)
Using polling for sockets would avoid setting up some threads,
especially the autocalibration threads in the Listener, i.e.
autocalibrate() and autocalibrate_hello(), and listen_for_aps() in the
Aggregator.
See also select_tut(2).
- Eliminate remaining exit() calls to avoid memory leaks on exit.
- Handle IPv6.
= Global – maybe someday =
- Use atexit() / on_exit() instead of goto to clean at exit
- Support string-based positioning requests
Along with the binary requests, implement a string-based request
format. Each packet will contain a CSV string describing the request's
information. This will make easier to develop clients in other
languages than C and over other protocols than UDP.
- Add a library to parse aggregation files, handling the different
formats ?
It could be used by Positioner and external programs.
= Aggregator =
- An empty positioner_host should turn off communication with the
positioning server.
- Option to overwrite the output file (ask for a confirmation!).
- Outputs should include all the timestamps
The CSV output file and the socket should include the capture
timestamp and the reception timestamp on the Aggregator. This would
help the Positioner's --replay option (to be adapted to use the
aggregation time instead of the request time).
- Allow a different aggregation timeout for each request type.
- Refactor add_captured_request().
- got_request(): option for the maximal difference time
For implicit packets, we consider that packets from the same MAC and
received within an interval of 10ms are part of the same request. We
should define an option to allow user to choose the time he wants.
= Listener =
- Known bugs:
- On OpenWrt Backfire (10.03), when the program is statically linked,
using both -D and -A at the same time makes the program block on
the first thread. This is a known issue of uClibc:
https://dev.openwrt.org/ticket/4809
- Better handling of retried packets ?
Currently we ignore any packet with the Retry bit. This can lead to
completely ignore the given packet (or even request) if the original
packet is never received, which can happen under certain
circumstances.
Before changing this behaviour, it must be thought seriously: taking
into account only the original packet means that all the listeners
measure the signal strength from the same packet, which is what we
want in most cases.
This could be implemented as an option.
- Refactor and review read_packet().
- Filter the number of implicit packets per mobile?
Currently, in passive or mixed mode, //all// the received packets are
transmitted to the aggregator.
- Option "read-only autocalibration" allowing to listen for
autocalibration requests without sending requests.
- read_packet(): use ieee80211_header_size for all implicit packets
Currently the size is corrected only for data packets.
= Client =
- Log sent requests?
- Qt version of the client
We could either modify the current OwlPS Client to add an optional GUI
(optional dependency on Qt) or create a new, purely graphical client.
A Qt client would be portable to Android thanks to the Necessitas
project: http://necessitas.kde.org/
A port of Qt for Apple iOS was also developed but seems inactive:
https://qt.gitorious.org/qt/qt-iphone-clone/activities
= UDP to HTTP =
- Delete inactive clients after a given amount of time.
- Refactor prepare_answer().
= Positioner =
== Known bugs ==
- With autocalibration on in meshing mode, all the CPs declared in
listeners.csv must be in coverage or the reference points will not be
generated.
== Algorithms ==
- Add to the result the information "area error" (whether on not the
computed result is in the same room as the real position).
- Handle the power and antenna gain difference between the calibration
mobile (or pseudo-mobile, for autocalibration) and the mobile to
position.
- MinMax: add option to set the step (and use a different step for X, Y
and Z?).
- MinMax: reduce the research space as proposed by F. Lassabe (PhD
report, Fig. 2.6 p. 38 and explanations pp. 38-39).
== Autocalibration ==
- Generate reference points in 3D.
- To select the reference CPs, we should use the distance along with the
angle and coverage.
- Handle 2 CPs, not only >2 CPs with mesh autocalibration.
- Option z-level-number: the Z coordinate represents a floor number. If
unactivated, Z is in metres and the topology is used to determine a
change of level.
- Generated CalibrationRequest have always nb_packets == 1.
It is tricky to set a number of packets for generated requests, since
the CPs can transmit a different number of packets in their
autocalibration requests. A solution could be to use the biggest
number amongst the reference autocalibration requests used. A simpler
solution could be to assume that all the CPs transmit the same number
of packets.
== User interface ==
- Move autocalibration-related options from the "positioning" category
to a dedicated category.
- When reading the CPs, add them to the mobiles' list (or another way to
be able to have a single entry for a CP).
- Add option positioning.self-calibrate (or autocalibrate), short option
-A, to activate automatically the options needed by the
autocalibration.
- Add option dump-configuration (displays the config & exits), if it is
possible with boost_program_options.
- Improve --verbose (and/or debug level): print the options, etc.
- Case-insensitive string comparison (for algorithm names, etc.).
- Use a prefix for configuration files (search for config files set with
relative path in owlps-positioner.cfg in the same directory).
== Miscellaneous new features ==
- Add a statistics framework (StatisticsMedium), that would provide
information at the end of the execution.
The information that could be provided include:
- number of requests treated, by type
- for each algorithm:
- mean error
- number of results for which the room is wrong
- number of results for which the floor is wrong
- error information by room
== Refactoring ==
- Move Stock::generate_reference_point() & friends to a dedicated class.
- Write a class for Request::type?
CalibrationRequest::direction uses a dedicated class Direction, why
not Request::type? That would simplify writing of the type to streams
(no need to cast each time any more).
- Wi-Fi devices' list
- Merge Stock::mobiles & Stock::aps?
- Factorise AccessPointsReaderCSV & MobilesReaderCSV?
- Members renaming
- InputMedium:
- current_line_nb & get_current_line_nb()
- get_next_request() > read_next_request()
- Input: get_next_request() > read_next_request()
- Area: p_min et p_max > coord_min et coord_max
== Optimisation & code improvement ==
- Eliminate const_cast where possible.
- Use the 'ws' stream manipulator to eat whitespaces.
- Change ss_t type to int_fast8_t?
- Multithreading (OpenMP?):
- algorithms (parallelize computation of the different algorithms for
the same request);
- and/or requests (parallelize computation of several requests).
- ReferencePoint: the request list should be an unordered_set instead of
a vector, to guarantee the unicity of the elements.
- Pre-allocate vectors' memory with reserve().
"C++ en action", p. 217.
- Copy of containers to streams (in some operator<<() for instance).
"C++ en action", p. 275.
- Review all the classes to respect principles exposed in "Coder
proprement", chapter 6, p. 103: do not implement accessors for each
class attribute, etc.
== Unit tests ==
- Classes with incomplete tests:
- InterlinkNetworks: compute()
- Measurement: average, variance, standard deviation, operations
- MinMax: trilaterate_2d()
- OutputCSV: write(ResultList)
- Point3D: distance & angle operations
- PosUtil: most functions
- ReferencePoint: operations
- Stock: a lot of functions
- Classes without tests or with mock test files (possibly untestable or
not worth testing, to be checked):
- Autocalibration
- AutocalibrationLine
- AutocalibrationMesh
- CapturePointsReaderCSV
- CartographyAlgorithm
- CSVStringReader
- FBCM
- FRBHMBasic
- Input
- InputDataReader
- InputLogMedium
- InputMedium
- InputUDPSocket
- MobilesReaderCSV
- NSS
- Output
- OutputMedium
- OutputNetworkSocket
- OutputTCPSocketEvAAL
- OutputUDPSocket
- Positioning
- PositioningAlgorithm
- ResultList
- TrilaterationAlgorithm
- TrilaterationMethod
- Timestamp: there is a probability of 10^-6 that the value in
nanoseconds and the rounded value in milliseconds are identical, in
which case one of the tests can fail. This could be fixed by retrying
the same test a couple of times.