School of Information Sciences Geoinformatics Laboratory
Automatic Selection of Landmarks for Navigation Guidance Rui Zhu Geoinformatics Laboratory School of Information Sciences University of Pittsburgh October, 2013
School of Information Sciences Geoinformatics Laboratory
Outline • Motivations • Methods • Feature Selection &Data Source • Preprocessing • Model & Experiments • Results &Conclusions • Future work
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Motivations 1. Landmarks play a major role in turn-by-turn instructions (citations) Landmarks=salient objects in the environment that aid the user in navigation and understanding the space (Sorrow & Hirtle, 1999) Techniques to select those salient objects http://sites.garmin.com/nuvi
Tourist attractions ≠ Landmarks
Automatic Selection of Landmarks for Navigation Guidance
https://maps.google.com/maps
POIs ≠ Landmarks
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Motivations 2. Artificial neural network • Simulate humans’ complex decision making processes • Deal with noisy and erroneous data
• People’s selection of landmarks is complex • The data used to select landmarks are always free spatial database, so data are noisy and erroneous
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory OSM
LiDAR
Social Network (Foursquare)
Width, Length, Area, Type
Method
Spatial Query
+
Sampling/ Satatistics
Ratios, number of adjacent segments, number of adjacent buildings
Height
+
+
API
number of checkins, number of total visits, number of user reviews
Step 1
Search Engine (Google, Bing Images)
API
number of images, number of search results
Step 2
Absolute Values Buffering (500m)
Label instances for model
Relative Values
Step 1
Randomly select 70% of the data as train data
Automatic Selection of Landmarks for Navigation Guidance
Step 2
Step 3
Build the artificial neural network model using the train data
Use the rest 30% instances to test data
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Feature Selection & Data Source • Static features: are those that place more emphasis on the ways in which humans observe geographic objects; individuals’ behaviors have no influence on static features. • Dynamic features: are based on the ways in which humans interact with geographic objects; people’s past behaviors are used to collect them
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Feature Selection & Data Source LiDAR (Height)
LatLon (intersection)
Google (Number of search results)
Name/ Address
OSM (Width; Length; Area; Three ratios; number of adjacent buildings; number of adjacent roads)
Name/ LatLon
Foursquare (Check-ins; Total visitors; Users review)
Name/Address
Bing (Number of images) Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Preprocessing - Conversion • Convert absolute values to relative: the attractiveness of the object should be a relative measure within a limited boundary (Raubal & Winter 2002; Tezuka & Tanaka 2005; Elias 2003). • In our dataset (Pittsburgh), we set the buffering range to 500 meters. •
Bad Examples
Good Examples
Scenario 1a
m
Density: low Buffering: large
: target object
Scenario 1b
n
Density: high Buffering: small
: adjacent objects
Automatic Selection of Landmarks for Navigation Guidance
Scenario 2a
Scenario 2b
m
Density: high Buffering: large
n
Density: low Buffering: small
m and n are the buffer ranges, and m>n Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Preprocessing-Labeling • Labeling objects for ANN to learn (1). Manual labeling Subjects familiar with Pittsburgh; Tools: Google Maps (2D visual and structure attractiveness) Images (detailed visual information)
(2). Rule-based labeling Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Preprocessing- Labeling- Rules • Five assumptions • Threshold
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Model & Experiments • Two hypothesis: H1: Introducing dynamic features could improve the performance of the model; H2: Applying rule-based labeling approach largely outperforms manual labeling. • Four experiments
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Model & Experiments • Structure of ANN Ø One-hidden-layer: One hidden layer can approximate any function that contains a continuous mapping from one finite space to another (Jeffheaton, 2008 ) Ø How many hidden nodes? Three rule-of-the-thumb:
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Results & Conclusions 1
The average gap is within 5% à Structures based on the three rule-of -thumb methods have no significant difference when building the model à Select Structure 3 as default structure Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Results & Conclusions 2 Presision
Static
Recall
Static+Dynamic
100.00%
100.00%
90.00%
90.00%
80.00%
80.00%
70.00%
70.00%
60.00%
60.00%
Static
Static+Dynamic
50.00%
50.00% Manual
F-Measure
Manual
Rule-based
Static
ROC Area
Static+Dynamic
100.00%
100.00%
90.00%
90.00%
80.00%
80.00%
70.00%
70.00%
60.00%
60.00%
50.00%
Rule-based
Static
Static+Dynamic
50.00% Manual
Rule-based
Manual
Rule-based
Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Results & Conclusions 2 • rule-based labeling significantly outperforms manual labeling by 10% in average, especially for precisionà Hypothesis 2 • both static and dynamic features combined is shown to be superior over using only static features, although the improvements, which are around 5% on average, are not as significant as anticipated à Hypothesis 1 Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014
School of Information Sciences Geoinformatics Laboratory
Future work • Improvement of labeling methods Rules: involve subjects into the design of the these rules Manual: encourage users to contribute and recommend landmarks using social networks
• Improvement of features Static: color; texture of the building (image processing) Dynamic: more features could be extracted from web resources
• Improvement of models More advanced models (like SVM) could be experimented in this work Automatic Selection of Landmarks for Navigation Guidance
Rui Zhu, March 5th 2014