A tomated Assessment of Tornado-Induced Automated Tornado Induced Ind ced Building B ilding Damage g Based on Laser Scanning g Alireza G. Ali G Kashani K h i ((alireza.geranmayeh@gmail.com) g y @g ) Ad i Advisor: Dr. D Andrew A d Graettinger G i g D p t Department t off Ci Civil, il, Construction, C t ti , and d Environmental E i t l Engineering, E gi i g, The Th University U i ity off Alabama Al b Introduction d
Research h Approach pp h
Results l
• Damage assessment after natural disasters is important for • Data processing algorithms forensic engineering and loss estimation. estimation
• Key is to preserve perishable data of damaged buildings. • 3D laser scanning is an effective technology that captures geometric information with high precision.
I Input GIS layers l
• Point cloud of damaged buildings
• Pre-tornado Pre tornado aerial image
• Dense 3D point cloud data produced by laser scanning virtually reconstructs the damage site and enables engineers to take measurements and retrieve geometric i f information. i LLoss off rooff covering Uplift of roof
Failure of large g section of roof
Estimating wind speed based on damage
D processing Data i algorithms l ih
• Filter noise points (ground, trees, etc.) • Identify roof and wall surfaces • Detect Damage • Quantify the extent of damage • Extract geometric information: • Area of roof covering defects • Percentage off sheathing h h lloss • Roof pitch • Spat Spatial a relationships e at o s ps bet between ee bu building d g
O Outputs
• Quantitative damage information
• GIS damage map • GIS G S wind i d speed d map
Correlation of wind speed & roof damage
Automatically generated wind speed map
Extracting i rooff planes l and d their h i spatial i l relationships l i hi to the h tornado and wind direction
surfaces and tornado/wind / direction
• Estimate wind speed based on damage
• Algorithm tests in controlled conditions Angle between the plane normal and tornado path
Failure of d garage door
3D data d t collection ll ti with ith llaser scanners
Failure of some exterior walls
Information extracted for the selected plane: Sheathing loss: 35% Roof pitch: 33 degree Distance: 132 meter Angle:18 Angle: 8 degree
Broken window
3D point i t cloud l d off damaged d d buildings b ildi
• Manually taking measurements and processing data is challenging and time consuming due to the large amount of data collected and the repetitive nature of manual operations and calculations calculations.
• This PhD research aims at developing a data processing
Simulated data
Damaged roof model
Laboratory scans
• Algorithm l i h tests iin reall conditions di i
Detectingg roof coveringg defects Red: Roof shingles Black: Plywood Blue: Roof felt
framework to automatically analyze tornado-induced tornado induced b ildi d building damage based b d on laser l scanning i data. d Source: Sou ce http://extremeplanet.me/ ttp //e t e ep a et e/
Source: Sou ce FEMA at http://fema.maps.arcgis.com/ ttp // e a aps a cg s co /
p – Tuscaloosa,, AL Residential complex
Practical Implications
• Performance data to improve structural analysis, analysis design, design and d construction i
Extracted roof planes and spatial information
g Residential neighborhood – Moore,, OK
Results
• Automatedd ddata processing i fframeworkk for: f Identifying roofs and walls and calculating percentage of loss
Detected roof defect regions
• Sensitivity S i i i analysis l i off framework f k parameter settings i
• Generation of damage maps for disaster response and recovery
• Loss database updating for management of insurance and disaster assistance programs
• Preliminary P li i l loss estimation ti ti ffor costt analysis l i off repair i and d reconstruction projects p j Automatically generated GIS damage maps
Influence of GIS raster cell size on accuracy of (a) roof detection and (b) wall detection
Information Extraction and Automated Reasoning for Automated Regulatory Compliance Checking in the Construction Domain Jiansong Zhang, jzhang70@illinois.edu, Advisor: Nora El-Gohary, University of Illinois at Urbana-Champaign 1. Problem Statement
4. Preliminary Results
Manual checking of construction regulatory compliance is timeconsuming, costly, and error-prone. Automating the process of compliance checking is expected to reduce time, cost and error of the process. Previous research and development efforts in automated compliance checking (ACC) have paved the way, but they have been limited because 1) they require manual understanding, extraction, and formulation of regulatory requirements, and 2) they lack the level of knowledge representation and reasoning that is needed for compliance analysis and checking.
4.1 Implementation Method
4.4 Automated Reasoning Results
Information Extraction
Information Transformation
Automated Reasoning
Implementation Method of the Main Algorithms
4.2 Information Extraction Methodology and Results
2. Research Objective Preprocessing
Feature Generation
Target Information Analysis
Tokenization Sentence Splitting Morphological Analysis
POS Tagging Phrase Structure Analysis Gazetteer Compiling Ontology-Based Semantic Analysis
Target Information Identification Extraction Sequence Resolution
Evaluation
Theoretically/empirically developing methodologies/algorithms and corresponding knowledge representations that 1) automatically extract regulatory and project information from textual regulatory documents and BIM models, respectively, 2) encode the information in a semantic format, and 3) automatically reason about the information for compliance assessment.
Information Representation
Partial Automated Compliance Checking Results of International Building Code 2006
(1) Project BIM Model
(2) Regulatory Documents
(5) IE Algorithms
(6) ITr Algorithms
(7) AR Algorithms
5.1 Scientific Contributions
Development of Information Extraction Rules Extraction Execution
3. Research Method Utilizing a theoretically-based, empirically-driven methodology that leverages knowledge and techniques in natural language processing (NLP), semantic modeling and automated reasoning (AR) to develop methodologies/algorithms and corresponding knowledge representations for automated information extraction (IE), automated information transformation (ITr), and automated compliance reasoning for regulatory information and project information in the construction domain.
5. Scientific Contributions and Practical Significance
Development of Extraction Rules for Extracting Single Information Elements (Matching Pattern Construction; Feature Selection; Semantic Mapping;) Development of Rules for Resolving Conflicts in Extraction
Proposed Information Extraction Methodology
Information Extraction Results for Chapter 19 of International Building Code 2009 Semantic Compliance Deontic Quantity Subject Quantitative Comparative Quantity Quantity Information Subject Checking Operator Unit/ Total Restriction Relation Relation Value Restriction Element Attribute Indicator Reference Precision 0.941 1.000 0.935 0.979 0.947 1.000 0.976 0.976 1.000 0.969 Recall
0.941
0.833
0.956
0.958
0.931
0.935
0.976
0.953
0.900
0.944
F-Measure 0.941
0.909
0.945
0.968
0.939
0.967
0.976
0.964
0.947
0.956
4.3 Information Transformation Results Information Transformation Results for Chapter 12 of International Building Code 2006 Performance Measures
(8) Extracting Project Information (Logic Facts)
(9) Extracting Rules (Logic Rules)
(10) Testing Compliance
(11) Reporting Compliance
Concept Logic Clause Elements
Relation Logic Clause Elements
(3) Semantic Model
(4) NLP Techniques
Automated Compliance Checking Framework
Total
Precision Recall F-Measure Precision Recall F-Measure Precision Recall F-Measure
Top-Down Method
Bottom-Up Method
0.961 0.923 0.941 0.948 0.927 0.938 0.954 0.925 0.939
0.997 0.976 0.986 0.938 0.951 0.944 0.962 0.961 0.962
• Offer efficient-to-develop, semantic rule-based NLP methods for IE and ITr that can help capture domain-specific meaning; • Combine domain knowledge and NLP knowledge to achieve deep NLP; • Pioneer work in utilizing semantically deep NLP approach in the construction domain; • Provide benchmark semantic methods/algorithms for IE and ITr in the construction domain; • Offer a new mechanism for processing and transforming complex regulatory text into logic clauses; • Pioneer work in utilizing logic programming (LP) for the automated reasoning functionality in ACC in the construction domain. 5.2 Practical Significance • Reduce the time and cost of compliance checking in the construction domain; • Improve the accuracy of compliance checking; • Support other applications of automated information processing in the construction domain.
Improving Construction Cost Escalation Estimation Using Macroeconomic, Energy and Construction Market Variables Mohsen Shahandashti1 and Baabak Ashuri2 1Ph.D.
Candidate, School of Building Construction, Georgia Institute of Technology, Email: sshahandashti3@gatech.edu 2Associate Professor, School of Building Construction, Georgia Institute of Technology, E-mail: baabak.ashuri@coa.gatech.edu
Introduction • Recently, the accuracy of construction cost estimates has been significantly affected by fluctuations in construction costs in the United States. • Construction cost fluctuations have been larger and less predictable than was typical in the past. • Cost escalation has become a major concern in all industry sectors, infrastructure, heavy industrial, light industrial, and building.
Research Approach
Summary of Results
Identification of leading indicators of construction cost variations
Leading Indicators of Construction Cost Variations: Macroeconomy
A pool of 16 candidate (potential) explanatory variables is initially selected based on a comprehensive literature review about construction cost variations.
200.0
190.0
Construction Cost Index (CCI)
150
240.0 190.0
6590
150.0
11
20
9
15
7
10
5
5
250.0
250.0
3
0
200.0
200.0
150.0
150.0
50
90.0
6550
40.0
6530
0
Producer Price Index
6510
Crude Oil Price
Unemployment Rate
90.0
50.0
40.0
0.0
Producer Price Index
Federal Funds Rate
250.0
8000
42
100.0
100.0
200.0
200.0
7000
40
50.0
50.0
6470
150.0
150.0
6000
0.0
0.0
6450
100.0
100.0
5000
38
50.0
50.0
4000
36
0.0
0.0
3000
34
6490
Consumer Price Index 15000
25
10000
Problems Related to Cost Variations
5000
Related Issues to Owner Organizations Related Issues to Contractors Hidden price contingencies Delayed or cancelled projects Inconsistency in budgets Unsteady flow of projects
• Bid loss due to cost overestimation • Profit loss due to cost underestimation
0
Dow Jones Industrial Average 25
15
150.0
6000.0
10
100.0
4000.0
5
50.0
2000.0
0
0.0
0.0
Prime Lending Rate 0.2000
Previous work in predictive modeling can be classified into two categories:
Number of Building 20000.00 Permits
120.000
15000.00
100.000
-0.0500
Inflation Rate
60.000 40.000
0.00
20.000
The explanatory variables of construction cost variations are identified from the pool of candidate explanatory variables using empirical tests including correlation tests, unit root tests, Granger causality tests, and Johansen’s cointegration tests.
Multivariate Time Series Modeling
Statistical Methods
Based on the results of Johansen’s cointegration tests, Vector Error Correction Models with various combinations of leading indicators and CCI are created. The long-run form of VEC model can be represented by the following equation:
• Current statistical methods, such as univariate time series models, do not have explanatory capability and they are just suitable for short-term forecasting (Wong and Ng 2010; Touran and Lopez 2006; Goh and Teo 2000; Pindyck and Rubinfeld 1998). • The forecasting power of causal models depends on the identification of appropriate leading variables (Wong and Ng 2010; Williams 1994). However, the leading variables of CCI are not known. • The temporal relations of variables is ignored in these models.
Number of Housing Starts
• Data from January 2009 to December 2011 is used to assess the predictability of the proposed multivariate time series model. • Data from January 1975 to December 2008 is used for conducting statistical tests and creating models. • The VEC model including CCI and COP is more accurate than those forecasted by the previously proposed univariate models (i.e., Seasonal ARIMA and Holt-Winters exponential smoothing).
Error Measures • Mean Absolute Percentage Error (MAPE)
p 1
yt Ai yt i Byt p C t i 1
Equivalent representation of the above equation:
yt 1 yt 1 ... p yt p C ut Ai ( I 1 ... i ),
Crude Oil Price
Validation
Causal methods, such as regression models and neural networks, use economic and market factors, such as Prime Lending Rate, to forecast future trends of CCI (Trost & Oberlender 2003; Williams 1994; Akintoye & Skitmore 1994; Koehn & Navvabi 1989).
Gaps in Knowledge
0
Gross Domestic Product GDP Implicit Price Deflator
Causal Methods
Statistical methods, such as time series analysis and curve fitting, forecast future trends of CCI based on past values (Ng et al. 2000; Wang & Mei 1998; Fellows 1991; Taylor & Bowen 1987) .
50
CCI 1.2301 - 0.7266 CCI - 0.1979 2.7593 CCI COP - 0.0118 1.6681 COP 0.0165 - 0.6801 COP t t 1 t 2 - 0.0943 - 1.0293 CCI 0.0104 - 1.0240 CCI - 0.0267 - 0.0384 COP 0.0348 0.0277 COP t 3 t 4 - 0.0678 1.5072 CCI 0.1189 - 1.2923 CCI 13.0903 u1 - 0.009 - 0.1340 COP - 0.0034 0.1599 COP 0.2164 u t 5 t 6 2 t
Money Supply
5000.00
100
Based on diagnosis tests, the multivariate time series model including Crude Oil Price (COP) and Construction Cost Index (CCI) is the best model among other combinations.
80.000
10000.00
0.0000
Average Hourly Earnings
Research Background
10000.0 8000.0
0.0500
10
Average Weekly Hours
200.0
0.1000
15
Employment level in 250.0 Cont.
20
0.1500
20
5
• Not enough bidders
Number of Housing Starts
Consumer Price Index
150
Number of Building Permits
250.0
Example of Construction Cost Variations
• • • •
25
140.0
6570
100.0
13
100
Energy Market
250.0
240.0
140.0 6610
Construction Market
for i 1,.., p 1
• Mean Square Error (MSE)
Notation: yt is the (K×1) vector of time series at period t K is the number of variables Ai (i=1,…,p-1) are (K×K) coefficient matrices of endogenous variables containing the cumulative long-run impacts
MAPE MSE
B is (K×K) coefficient matrix C is (K×1) vector of constants
VECM including CCI and COP 0.96% 10544.9
Seasonal ARIMA 1.40% 17921.6
Holts Winter-Exponential Smoothing 2.68% 86890.7
εt is (K×1) vector of error terms
B ( I 1 ... p ) The coefficients of these VEC models are estimated using Gaussian maximum likelihood procedure (Johansen 1995).
Research Objective
Diagnosis of models
The objective of this research is to create multivariate time series models for improving the accuracy of construction cost escalation estimation through utilizing information available from several indicators of macroeconomic condition, energy price and construction market.
• Lack of serial correlation among residuals is diagnosed using Breusch–Godfrey serial correlation Lagrange multiplier test (Breusch 1978; Godfrey 1978) • Constant variance of residuals is diagnosed using ARCH test (Engle 1982)
© 2014, Economics of the Sustainable Built Environment (ESBE) Lab
Conclusions
•
•
Two macroeconomic variables (consumer price index and producer price index), two construction market variables (number of building permits and number of housing starts), and an energy market variable (crude Oil price) are helpful for predicting construction cost variations ahead of time. The multivariate time series model including crude oil price performed the best on predicting construction cost variations ahead of time.
Development of an Automated Progress Monitoring and Control System P.hD. Student: Reza Maalek (rmaalek@ucalgary.ca); Supervisors: Dr. Janaka Ruwanpura, Dr. Derek Lichti
1. Introduction/ Problem Statement TQM
a) Monitoring in practice
Construction Projects
(Golparvar-Fard et al. 2009a, Golparvar-Fard et al. 2011a, Maalek et al. 2014)
Manual
Time Consuming To Labor Intensive Justify Prone to Errors
Non-systematic Untimely identification of the causes of delays and cost overruns
Collection of Limited data
Inaccurate Performance Measure Increase chances of Claims Increase chances of Rework
b) Controlling in practice: (Golparvar-Fard et al. 2011b, McCullouch, B. 1997)
Highly dependent on the accuracy of the Monitoring stage 30-50% of site supervisor’s working time spent controlling
If Improved
Allocating More Safety Time to Improve Productivity and Communication
2. Background /Literature Review a) State-Of-the-Art Automated Monitoring: (Maalek et al. 2014) 3D coordinates of Structural Elements
Identification of “Scope of Work Performed”
b) How LiDAR works?
Camera (Photogrammetric Bundle Adjustment) 3D Range Imaging LiDAR
LiDAR Point Clouds
Object of Desire
3. Objectives and Methodology: Developing a “Fully Automated” Progress Monitoring and Control System comprised of the following stages: Feature Extraction
Monitoring
Controlling Scope of Work Performed
LiDAR
As-built Model Generation
Compare to As-planned
Addressing Three Major Concerns:
On Schedule
Ahead of Schedule
Behind Schedule
Rework due to poor quality
Scan #1 Scan observing the Maximum # of Facets: Assume Scan #n
1. Minimum # of Scan-stations:
Scan-station 1=Scan #n
?
Simulate LiDAR Point Clouds for the Scan-stations
:
Scan #n
As-Planned Model
Erin Dvorak
New As-planned Model
2. Automated Feature Extraction: Until a “Point Cloud” is generated for each Facet
As-built Model
?
:
As-built Point Cloud
Robustly Estimate Principal the Covariance Components
Define a neighborhood for each point
Eigenvectors Eigenvalues Smoothness Continuity Surface Shape
3. Automate Registration of As-built Model:
(X,Y,Z) (ω,φ,κ)
:
Minimum Requirements for Rigid body Transformation
As-planned Model
As-built Model
4. Experiments & Results:
Two convergent linear features (Preferably Orthogonal lines)
Three non-collinear points
Initial ICP Rigidbody Registration Transformation EOPs
Assign Points to Structural Elements Until EOP Stable
Closest points in the Planned and As-built Perform Rigidbody Transformation
Translation: (X,Y,Z) Rotation: (ω,φ,κ)
5. Contributions and Significance: Development of an automated Monitoring and Control System Construction Industry Benefits: 9 cm Point to Point a. Reduction of the project managers’ travel time (Giretti et al. 2009) b. Improving time, cost and quality of onsite data (Golparvar-Fard et al. 2009a) c. Reduction of preparation of Progress Reports (Golparvar-Fard et al. 2009b) Wall a) 2.4 cm Vertical d. Timely identification of work progress deviations (Maalek et al. 2014) Dimension b) 7.5 cm Horizontal e. Reducing Rework due to poor quality (Golparvar-Fard et al. 2011a) Compliance f. Reducing construction claims (Semple et al. 1994) g. Structural stability monitoring and control (Park et al. 2007) MRSE Accuracy
As-Built Generation 26 Walls 22 in the LOS 9.4
13.5
As-Planned Model
As-Built Model
Quantifying Human Mobility Perturbation
Qi Ryan Wang wangqi@vt.edu Advisor: Prof. John E. Taylor
Under the Influence of Tropical Cyclones BACKGROUND
RESULTS
• Tropical cyclones are becoming more frequent1,2 and intensified (2%-80%)3. Number of Trips
Frequency
• Hurricane Sandy: 202 deaths
(i) Travel
• Several events have caused significant fatalities, injuries, and damage in recent years.
Sandy (Cat. 3) in NYC r1=1-100m r2=100-500m r3=500m-1km
Oct. 29, 2012
• Typhoon Haiyan: 6000+ deaths
• How can human mobility perturbation be quantified?
Data Sample
Figure 1.
DATA COLLECTION
Oct. 29, 2012
Haiyan (Cat. 5) in Tacloban r1=1-100m r2=100-500m r3=500m-1km
r4=1-5km r5=5-10km r6=>10km
Other days
Average value from Oct. 30 to Nov.9, 2012
Nov. 5-7, 2013
Nov. 8-10, 2013
Average value from Nov. 11 to Nov. 28, 2013
Range of Δr
r4=1-5km r5=5-10km r6=>10km
Other days
Range of Δr
P(Δr )
Patterns
Oct. 29, 2012
Other days
Average value from Oct. 30 to Nov.9, 2012
Quantification
• Do tropical cyclones change human mobility patterns?
(ii) Movement
• Do tropical cyclones perturb human travel frequencies?
r1=1-100m r2=100-500m r3=500m-1km
Range of Δr
(iii) Perturb.
KEY QUESTIONS
Other days
Average value from Oct. 30 to Nov.9, 2012
• Typhoon Wipha: 39 deaths • Human mobility plays a key role in disaster evacuation, response and relief.
r4=1-5km r5=5-10km r6=>10km
Wipha (Cat. 4) in Tokyo
Δr(/m)
JSDS = 0.022 TS = ~24 hr
Oct. 29, 2012
Other days
Average value from Oct. 30 to Nov.9, 2012
Δr(/m)
JSDW = 0.063 TW = ~48 hr
Nov. 5-7, 2013
Nov. 8-10, 2013
Average value from Nov. 11 to Nov. 28, 2013
Other days
Δr(/m)
JSDH = 0.162 TH > 3 weeks
Figure 2. (i) Short-distance trips increased while long-distance trips reduced in Sandy and Wipha, but the changes are more chaotic during Haiyan. (ii) Movement data fit to truncated power-law distribution well for all three tropical cyclones; (iii) Quantifying both strength and duration of human mobility perturbation.
CONCLUSIONS • Tropical cyclones can cause human movement perturbation4,5. • Power-law still governs human mobility6,7 unless the storms are extremely powerful. • Quantification shows categories are associated with perturbation strength and duration. 4pm Oct. 29 to 4pm Oct. 30
4pm Oct. 29 to 4pm Oct. 30
REFERENCES: 1. Ferris, E., D. Petz, and C. Stark, The Year of Recurring Disaster: A Review of Natural Disasters in 2012, 2013, The Brokings Institution. 2. Landsea, C.W., et al., Impact of Duration Thresholds on Atlantic Tropical Cyclone Counts*. Journal of Climate, 2010. 23(10): p. 2508-2519. 3. Knutson, T.R., et al., Tropical cyclones and climate change. Nature Geoscience, 2010. 3(3): p. 157-163. 4. Bagrow, J.P., D. Wang, and A.-L. Barabasi, Collective response of human populations to large-scale emergencies. PloS one, 2011. 6(3): p. e17680. 5. Horanont, T., et al., Weather Effects on the Patterns of People's Everyday Activities: A Study Using GPS Traces of Mobile Phone Users. PloS one, 2013. 8(12): p. e81153. 6. Brockmann, D., L. Hufnagel, and T. Geisel, The scaling laws of human travel. Nature, 2006. 439(7075): p. 462-465. 7. Gonzalez, M.C., C.A. Hidalgo, and A.-L. Barabasi, Understanding individual human mobility patterns. Nature, 2008. 453(7196): p. 779-782.
Vahid Balali
Segmentation and Recognition of Roadway Assets from Car-Mounted Camera Video Streams Using a Scalable Non-Parametric Image Parsing Method
(c)
(b)
(d)

Given a query (test) image Retrieval set of 200 similar images by taking the minimum per-feature rank of four global features

For each test superpixel Xi described by multiple local feature đ?‘“đ?‘–đ?‘˜ , compute a likelihood ratio score for each class a found in the retrieval set đ?‘ƒ(đ?‘‹đ?‘– |đ?‘Ž) L đ?‘‹đ?‘– , đ?‘Ž = đ?‘™đ?‘œđ?‘” = đ?‘™đ?‘œđ?‘” đ?‘ƒ(đ?‘‹đ?‘– |đ?‘Ž)
Create and validate an automated roadway asset data collection and assessment tool  Data Collection, analysis, and representation  A new scalable method for segmentation and recognition of roadway assets from videos  Reduce the time and effort for road inventories especially for guardrails and light poles
 
Semantic Labels
Asset Labels
Geometric Labels
Graph-based Segmented Superpixels
Guardrail
Tree
Pavemen t
Shoulder
Pole
Traffic Sign
Query Video Frame
(d)
Data
5. Data Collection and Setup 
Smart Road Dataset

Interstate I-57 Dataset
450 400 350 300 250 200 150 100 50 0
Semantic Labels
Traffic Sign
Guardrail Pavement Marking Asphalt Pavement
Vertical
Sky
Horizontal
Testing Image Geometric
Ground Truth Geometric Labels
Asphalt Pavement
Concrete Pavement
Guardrail
Horizontal
Vertical
Sky
Pole
Ground Truth Semantic Labels
Label Superpixels
Extract Superpixel Features
Query Frame Superpixels
Tree
Ground Truth Segmentation
2500 2000 1500 1000 500 0
Segmented Video Frames
Pavement Marking
Average accuracy of 78.67 % Smart Road Dataset
Testing Image Semantic
Traffic Signal
0 1
Likelihood Ratio Scores per Class for Each Superpixel
Bank of Superpixel Filters
Process
Video Frames
Geometric Labels 800 700 600 500 400 300 200 100 0
Query
đ??¸đ?‘ đ?‘šđ?‘œđ?‘œđ?‘Ąâ„Ž (đ?‘Žđ?‘– , đ?‘Žđ?‘— ) đ?‘‹đ?‘– ,đ?‘‹đ?‘– ∈đ?œ“
đ?œ =
Retrieval Set of Training Images
6. Experimental Results
2500 2250 2000 1750 1500 1250 1000 750 500 250 0
Bridge
Bank of Global Filters Detect Global Features
Pavement Marking
(b)
đ?œ (đ?‘Žđ?‘– , đ?‘”đ?‘– ) đ?‘‹đ?‘– ∈đ?œ“
(e)
(c)
đ?‘˜
đ?‘ƒ(đ?‘“đ?‘–đ?‘˜ |đ?‘Ž) đ?‘™đ?‘œđ?‘” đ?‘ƒ(đ?‘“đ?‘–đ?‘˜ |đ?‘Ž)
Simultaneously solve for a ďŹ eld of semantic and geometric labels over the image by optimizing
Training Video Frames
(a)
đ??¸đ?‘‘đ?‘Žđ?‘Ąđ?‘Ž đ?‘‹đ?‘– , đ?‘Žđ?‘– + đ?œ† đ?‘‹đ?‘– ∈đ?œ‘
đ??ť đ?‘Ž, đ?‘” = đ??˝ đ?‘Ž + đ??˝ đ?‘” + đ?œ‡
Retrieval Image Set
đ?‘˜
đ?‘ƒ(đ?‘“đ?‘–đ?‘˜ |đ?‘Ž) = đ?‘ƒ(đ?‘“đ?‘–đ?‘˜ |đ?‘Ž)
Use MRF inference to solve for the label field a = đ?‘Žđ?‘– over the entire test image đ??˝ đ?‘Ž =
3. Graphical Overview Query Video Frame

đ?‘–đ?‘“ đ?‘Ąâ„Žđ?‘’ đ?‘Žđ?‘– đ?‘–đ?‘ đ?‘œđ?‘“ đ?‘Ąâ„Žđ?‘’ đ?‘Ąđ?‘Śđ?‘?đ?‘’ (đ?‘”đ?‘– đ?‘‚đ?‘Ąâ„Žđ?‘’đ?‘&#x;đ?‘¤đ?‘–đ?‘ đ?‘’
Markov Random Field
Semantic Labeling Geometric Labeling
Asset Label
Average accuracy of 82.02 % Interstate I-57 Dataset
Soil
7. Contributions and Significance  
Video-based parsing and segmentation
   
Leveraging motion cues & temporal consistency
Improving the performance of 3D assets recognition (e.g. guardrails, light poles)
Reducing time and effort required for inventory An average accuracy of 88.24% for recognition An average accuracy of 82.02% for segmentation
Real-Time and Automated Monitoring and Control (RAAMAC) Lab
4. Method
1. Motivation  Manual data collection, time-consuming, subjective (a)  Significant expansion in size and complexity  Lack of techniques to easily track, analyze and visualize 2. Research Objective and Highlights
University of Illinois at Urbana-Champaign Advisor: Mani Golparvar-Fard