Urban Building Usage Labeling by Geometric and Context Analyses of - PowerPoint PPT Presentation

Institute of Cartography and Geoinformatics | Leibniz Universität Hannover Urban Building Usage Labeling by Geometric and Context Analyses of the Footprint Data Hai Huang, Birgit Kieler and Monika Sester

Introduction The Usage (use and occupancy) information of buildings ►  is of great interest for, e.g., navigation, city planning, emergency management, etc.  is not available in a consistent way in the volunteered data  is not always available even in the official cadastral maps Our approach to automatic labeling: ►  Pre-defined category: residential (family house, apartment building), com m ercial (shopping mall, office building), industrial (factory, warehouse) and public (school, hospital, theater, etc.)  High-level local geom etric features and contextual constraints  Markov Random Field (MRF) for the modeling of the building netw ork Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 2

Local features Influence of the local geometric features ►  Footprint area • (+ ) distinguish e.g., industrial and single-family house • (-) hard to separate public and industrial building without considering the shape  Length/ width ratio • (+ ) distinguish e.g., bar shape (often industrial, residential) and square shape (often public and commercial) • (-) only works for simple shapes (or using bounding box), cannot represent complex buildings  Higher level features are required to integrate multiple geometric attributes and provide more precise description to the footprint shape. Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 3

Local features Effective Width (EW) ►  Definition: the average width of the footprint along the centerline of the footprint ~ A B : building area L B : actual building length : length of the modified center line (red) Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 4

Local features Effective Width (EW) ►  Meaning: the general living/ movement space inside the building!  … despite the building complexity Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 5

Local features Branching degree (BD) ►  Definition: a score of the number and distribution of the building segments (the longest: “trunk”; the others: “branches”)  Basis: conventional straight skeleton Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 6

Local features Branching degree (BD) ►  Meaning: a measurement of the building complexity Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 7

Local features The local energy ►  EW-BD distribution  Each building can be represented as a point in this parameter space.  The class centroids are empirically given with generic values. Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 8

Local features The local energy ►  The probability that this building belongs to one of the classes is inversely proportional to its (standardized) distance to the centroid of the class.  Energy: a quaternary value of the probabilities (sum= 1) I.e., the probabilities that this building is supposed to be labeled to all the individual classes.  E.g., implies the building will be more likely labeled as “industrial” considering only the local features. Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 9

Context model The neighborhood relationship ►  Voronoi cells as basis Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 10

Context model Markov Random Field (MRF) ►  Vertices: centroids of individual buildings  Edges: neighborhood relationship Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 11

Context model The total energy ►  Unary term N R C I P  Binary term R 1 0 -1 0.5 C . 1 0 0.5 I . . 1 -1 P . . . 1 Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 12

Context model A global optimization ►  The configuration K, which leads to the maximum total energy  Four possible labels for each building/ vertices  Change of each label leads to a new configuration  Urban area  highly connected network  Computational intractable for direct solution… Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 13

Stochastic search Initialization M 0 current M n Propose a new state M’ Sample lables for the buildings Calculate the overall energy Accept the new proposal (Metropolis-Hastings criterion) Y N M n+ 1 = M’ M n+ 1 = M n Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 14

Experiments Dataset 1 (OSM): 94 buildings, Boston, United States ► Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 15

Experiments Ground truth ► Residential Commercial Industrial Public Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 16

Experiments Labeling using only local features (Accuracy= 72.3% ) ► Incorrect classifications Residential Commercial Industrial Public Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 17

Experiments Final results using both local and contextual information ► (Accuracy= 97.8% ) Incorrect classifications Residential Commercial Industrial Public Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 18

Experiments Dataset 2 (cadastral map): 456 buildings, Hannover, Germany ► (Accuracy= 89.7% ) Ground truth Labeling result Residential Commercial Industrial Public Incorrect classifications Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 19

Conclusion A novel full automatic labeling of building types (use and occupancy) ► Solely based on the building footprint data ► Four-classes category: residential, commercial, industrial and public ► Two new high-level geometric local features: effective width and ► branching degree MRF for the modeling of building network, integrating contextual ► constraints Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 20

Problems & outlook A general, rather rough category ► … against the wide variety of building types and definitions ► Some building types, e.g., educational, high-hazard, can actually ► NOT be derived solely from footprint data. Finer contextual knowledge can be integrated. E.g., an identified ► market square may increase the likelihood that the surrounding buildings are commercial!  New local features as well as contextual knowledge can be easily added in the proposed framework… for a refined or specialized classification. A general classifier is used all the datasets. In the future work, the ► local/ specified parameters of the classifier can be learned for individual cities.  A “spectral signature” of the city… Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 21

Thank you very much for your attention! Building Usage Labeling | ICC Dresden, 29 th Aug. 2013 Hai Huang | 22