|
@@ -261,7 +261,7 @@
|
261
|
261
|
booktitle = {Proc. IEEE Workshop on Motion and Video Computing WMVC '07},
|
262
|
262
|
year = {2007},
|
263
|
263
|
pages = {12--12},
|
264
|
|
- month = {Feb. },
|
|
264
|
+ month = {Feb. },
|
265
|
265
|
doi = {10.1109/WMVC.2007.4},
|
266
|
266
|
owner = {chris},
|
267
|
267
|
timestamp = {2009.12.03}
|
|
@@ -327,108 +327,6 @@
|
327
|
327
|
citeulike-article-id = {3480910},
|
328
|
328
|
citeulike-linkout-0 = {http://portal.acm.org/citation.cfm?id=1229555.1229562},
|
329
|
329
|
citeulike-linkout-1 = {http://dx.doi.org/10.1177/0278364907073775},
|
330
|
|
- comment = {NOTE: These notes were supplemented by looking at Liao's dissertation
|
331
|
|
- writeup of the same work.
|
332
|
|
-
|
333
|
|
- Overview: This is a supervised learning scheme done in several stages...
|
334
|
|
- the results from early stages are used to build the structure of
|
335
|
|
- the model in subsequent stages. Training is done for 3 people, and
|
336
|
|
- tested on a 4th person. For the test person, only GPS pings are used
|
337
|
|
- as input.
|
338
|
|
-
|
339
|
|
- specifics: * GPS traces from 4 people, 6 days per person * \~{}40,000
|
340
|
|
- GPS measurements per person * manually labeled all activities and
|
341
|
|
- significant places in these traces * used maximum pseudo-likelihood
|
342
|
|
- for learning
|
343
|
|
-
|
344
|
|
- hierarchical graphical model with 3 levels. top level: significant
|
345
|
|
- places (e.g. home, work, bus stop, parking lot, friend) midlevel:
|
346
|
|
- activity sequence (eg. walk, drive, visit, sleep, get on bus, pickup),
|
347
|
|
- bottom level: GPS trace (association to streeet map)
|
348
|
|
-
|
349
|
|
- * their typical GPS trace consists of approximately one GPS reading
|
350
|
|
- per second * GPS readings are segmented spatially; we have one activity
|
351
|
|
- node for each spatial segment of GPS pings ** e.g. a 12 hour stay
|
352
|
|
- at a single location is represented by a single activity node * if
|
353
|
|
- street map is available, segmentation is done in association with
|
354
|
|
- the street map (with 10m discretization) * "model can reason explicitly
|
355
|
|
- about the duration of a stay, for which dynamic models such as standard
|
356
|
|
- DBNs or HMMs have only limited support" [12,32]
|
357
|
|
-
|
358
|
|
- * two main groups of activities: navigation activities, and significant
|
359
|
|
- activities (in a single place, or at a transportation mode switch)
|
360
|
|
- * "to determine activities, our model relies heavily on temporal
|
361
|
|
- features, such as duration or time of day, and geographic information,
|
362
|
|
- such as locations of restaurants, stores, and bus stops." * "significant
|
363
|
|
- places are those locations that play a significnat role in the activities
|
364
|
|
- of a person: hoe, work, bus stops, parking lots typically used, homes
|
365
|
|
- of friends, etc...) model allows different activites to occur at
|
366
|
|
- same significant place; also, a significnat place can comprise mutiple
|
367
|
|
- different locations, mostly because of signal loss and GPS readings
|
368
|
|
-
|
369
|
|
- The CRF Model for Activity Recognition: 1) CRF for GPS denoising *
|
370
|
|
- break up street segments into 10 meter patches * measurement clique
|
371
|
|
- between each ping and all nearby street patches (Gaussian noise model
|
372
|
|
- from center of patch) * consistency cliques: put a Gaussian noise
|
373
|
|
- model on the \_difference\_ between the GPS displacement and the
|
374
|
|
- paired street patch displacement for consecutive pings * smoothness
|
375
|
|
- cliques: encourage street patch predictions to be stay on the same
|
376
|
|
- street, going in the same direction their conditional model for street
|
377
|
|
- patches given GPS pings is given in Eqn 26, bottom of page 10
|
378
|
|
-
|
379
|
|
- Output of the CRF model is taken as the spatial segmentation of the
|
380
|
|
- GPS pings... Hierarchical CRF is based on the segmentation, since
|
381
|
|
- one 'activity sequence' is associated with each sequence of points
|
382
|
|
- on the same 10m street patch * bottom layer of 3 layer CRF, now it's
|
383
|
|
- "local evidence" including: ** temporal information (e.g. time od
|
384
|
|
- day, day of week, duration of stay -- these can be discretized) clique
|
385
|
|
- functions are all binary indicators, one for every possible combination
|
386
|
|
- of temporal feature and activity ** average speed through segment
|
387
|
|
- -- discretized ** information extracted from geographic databses,
|
388
|
|
- such as whether a patch is on a bus route, close to abus stop, near
|
389
|
|
- a restaurant or grovery store; use indicator functions to model this
|
390
|
|
- information ** each activity node connected to its neighbors; e.g.
|
391
|
|
- extremely unlikely tha ta person will get on bus at one location
|
392
|
|
- and drive a car at neighboing location right afterwards (?)
|
393
|
|
-
|
394
|
|
- * Preliminary CRF\_0: just has activity nodes and local evidence nodes
|
395
|
|
- [notes from thesis] ** adjacent activity nodes are connected, and
|
396
|
|
- each activity node is connected to each 'ftr' with a pairwise potential
|
397
|
|
- ** features are *** gps based: streetpatch, timeofday (first timestamp,
|
398
|
|
- discretized into Moring, Noon, AFternoon, Evening, Night), dayofweek,
|
399
|
|
- duration, *** geo database: nearrestaurant, nearstore, nearbusstop,
|
400
|
|
- onbusroute [latter extracted from geographic databaddses] *** average
|
401
|
|
- speed thru segment (discretized - 'to allow multimodality')
|
402
|
|
-
|
403
|
|
- * Significant Places [from thesis] ** from the MAP activity sequence,
|
404
|
|
- they then (by hand? they refer to an IsSignificant() function) associate
|
405
|
|
- each activity with whether or not it belongs to a significant place
|
406
|
|
- (e.g. transport does not connect to a significant place, while getting
|
407
|
|
- on / off a bus does ...); * spatial clustering performed on the locations
|
408
|
|
- of significant places * the cluster centers are the 'signiciant plces'
|
409
|
|
- * edge between each place label and each activity label in its vicinity
|
410
|
|
- * NOTE: the 'place nodes' are not dynamic; we find K lat,longs that
|
411
|
|
- are the locations of the significant places; what we still ned to
|
412
|
|
- infer are the labels for these significant places; e.g. several of
|
413
|
|
- the lat,longs can be 'shopping', or 'friend's house, etc... certain
|
414
|
|
- place labels are associated with certain types of activities * they
|
415
|
|
- want to also add features that count the number of 'home', 'workplace',
|
416
|
|
- etc., labels that are used. However, this requires making a clique
|
417
|
|
- containing all the places, which can be quite large...
|
418
|
|
-
|
419
|
|
- generate\_places algorithm: * really not clear -- clustering places
|
420
|
|
- associated with the same activity? how can you get any confidence
|
421
|
|
- in the activities?
|
422
|
|
-
|
423
|
|
- * Each place (node at the top level of 3 level CRF) connects to all
|
424
|
|
- activities that seem to occur in the same place -- built into the
|
425
|
|
- structure of the model * multiple activities may occur at the same
|
426
|
|
- place
|
427
|
|
-
|
428
|
|
- * each GPS associated with a 10m patch on a street edge * [20] inference
|
429
|
|
- using loopy belief propagation, and parameter learning using pseudo-likelihoodreferences
|
430
|
|
- to chase down: bennewitz [4] learn different motion paths between
|
431
|
|
- places, [23] how to figure out types of places;},
|
432
|
330
|
doi = {10.1177/0278364907073775},
|
433
|
331
|
file = {Liao2007.pdf:Liao2007.pdf:PDF},
|
434
|
332
|
issn = {0278-3649},
|
|
@@ -661,7 +559,6 @@
|
661
|
559
|
algorithm in ...},
|
662
|
560
|
citeulike-article-id = {6212085},
|
663
|
561
|
citeulike-linkout-0 = {http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.7.8772},
|
664
|
|
- comment = {- history, algorithm, description of two bounds},
|
665
|
562
|
file = {Schapire1999.pdf:Schapire1999.pdf:PDF},
|
666
|
563
|
keywords = {boosting},
|
667
|
564
|
owner = {chris},
|
|
@@ -860,3 +757,17 @@
|
860
|
757
|
|
861
|
758
|
@comment{jabref-meta: selector_keywords:}
|
862
|
759
|
|
|
760
|
+@comment{jabref-meta: groupsversion:3;}
|
|
761
|
+
|
|
762
|
+@comment{jabref-meta: groupstree:
|
|
763
|
+0 AllEntriesGroup:;
|
|
764
|
+1 ExplicitGroup:Unread\;0\;Abowd1997\;Bellavista2008\;Bellotti2008\;Ca
|
|
765
|
+ros2005\;Davies2008\;Dornbush2005\;Eagle2004\;Floreen2008\;Garakani200
|
|
766
|
+9\;Hein2008\;Lee2007\;Lester2006\;Liao2007\;Liao2007a\;Lukowicz2002\;M
|
|
767
|
+ahdaviani2007\;Maurer2006\;Miluzzo2009\;Nicolai2006\;Parkka2006\;Patte
|
|
768
|
+rson2004\;Philipose2003\;Philipose2003a\;Philipose2004\;Rudstroem2004\
|
|
769
|
+;Schapire1999\;Schilit1994\;Schmidt1999\;Schmidt2008\;Serugendo2008\;S
|
|
770
|
+hen2004\;Song2005\;Stiefmeier2008\;Tentori2008\;Voida2002\;Wang2009\;W
|
|
771
|
+yatt2007\;Yang2008\;;
|
|
772
|
+}
|
|
773
|
+
|