1
Crowdsourcing the Acquisition and Analysis of Mobile Videos for Disaster Response Hien To, Seon Ho Kim, and Cyrus Shahabi Integrated Media Systems Center University of Southern California, Los Angeles, CA BigData in Social Media IMSC Retreat 2016 Introduction Popularity of mobile video, e.g., YouTube, Periscope Can be used to assess disaster damage across large geographical area Any way to effectively crowdsource mobile videos for disaster response? Motivation Obtain real-time information to enhance situational awareness during and after disasters Prioritize data collection with limited transmission Analyze collected data in timely manner Challenges Disrupted communication networks Spatial metadata System Architecture MediaQ for data collection – GeoQ for data analysis References Hien To, Seon Ho Kim, and Cyrus Shahabi. Effectively Crowdsourcing the Acquisition and Analysis of Visual Data for Disaster Response, In proceeding of 2015 IEEE International Conference on Big Data Proposed Solution Visual awareness (VA) maximization Selects videos (or key frames) that maximize total VA without exceeding a constrained bandwidth This problem is reducible to 0-1 Knapsack (or Maximum Coverage Problem) Assign videos to analysts GeoQ uses of grid-based partition Some analysts may be overloaded We use Kd-tree-based partitioning algorithm Each analyst handles equal number of videos Conclusion Proposed crowdsourcing framework for time- sensitive acquisition and analysis of video data Quantified visual awareness of a particular video or frame and introduced more effective spatial decomposition to assign videos to analysts Introduced two variants visual awareness maximization problem: uploading the entire videos and individual frames p R Field Of View (FOV) p: camera location d: camera direction : viewable angle R: viewable distance t: timestamp Haiti Earthquake, 70% cell towers were destroyed Ushahidi NGA GeoQ AI for DR Planetary Response Network USC MediaQ 0 0 0 1 3 2 5 2 0 0 4. Evaluate work cells 3. Assign videos to analysts 1.Crowdsour ce videos from community Urgency Map Analyst Serve r 2. Upload relevant videos Coverage Map Space Decomposition Video content Video metadata f 1 f 2 f 3 Video location Work cell A video has high VA if Enclosing work cell is urgent Video coverage is large 3-seconds video Results Synthetic Datasets [Ay et. Al. SIGSPATIAL’2010] Visual Awareness Maximization (Uniform Dataset) Statistics Value Spatial distributions Uniform, Gaussian and Zipfian #videos 1000 with varying length Region size 10 × 10 square km in Los Angeles Video Level 16 25 36 49 64 0 0.5 1 1.5 2 2.5 3 3.5 Grid Quadtr ee Analyst count Visual Awareness Kd-tree increases visual awareness by 300% 16 25 36 49 64 0 20 40 60 80 100 120 Grid Quadtr ee Analyst count Visual Awareness VA is 10 times higher in comparison to video-level Frame Level

Crowdsourcing the Acquisition and Analysis of Mobile Videos for Disaster Response

Embed Size (px)

Citation preview

Page 1: Crowdsourcing the Acquisition and Analysis of Mobile Videos for Disaster Response

Crowdsourcing the Acquisition and Analysis of Mobile Videos for Disaster Response

Hien To, Seon Ho Kim, and Cyrus ShahabiIntegrated Media Systems Center

University of Southern California, Los Angeles, CA

BigData in Social Media

IMSC Retreat 2016

Introduction Popularity of mobile video, e.g., YouTube,

Periscope Can be used to assess disaster damage across

large geographical area Any way to effectively crowdsource mobile videos

for disaster response?

Motivation Obtain real-time information to enhance situational

awareness during and after disasters Prioritize data collection with limited transmission Analyze collected data in timely manner

Challenges• Disrupted communication networks

Spatial metadata

System ArchitectureMediaQ for data collection – GeoQ for data analysis

ReferencesHien To, Seon Ho Kim, and Cyrus Shahabi. Effectively Crowdsourcing the Acquisition and Analysis of Visual Data for Disaster Response, In proceeding of 2015 IEEE International Conference on Big Data

Proposed SolutionVisual awareness (VA) maximization• Selects videos (or key frames) that maximize total VA

without exceeding a constrained bandwidth• This problem is reducible to 0-1 Knapsack (or Maximum

Coverage Problem)

Assign videos to analysts• GeoQ uses of grid-based partition

Some analysts may be overloaded • We use Kd-tree-based partitioning algorithm

Each analyst handles equal number of videos

ConclusionProposed crowdsourcing framework for time-sensitive

acquisition and analysis of video dataQuantified visual awareness of a particular video or frame

and introduced more effective spatial decomposition to assign videos to analysts

Introduced two variants visual awareness maximization problem: uploading the entire videos and individual frames

�⃗��⃗�

pR𝜃

Field Of View (FOV)p: camera locationd: camera direction: viewable angleR: viewable distancet: timestamp

Haiti Earthquake, 70% cell towers were

destroyed

Ushahidi

NGA GeoQ

AI for DR

Planetary Response NetworkUSC MediaQ

0 00

1 3

2 52

0 0

4. Evaluate work cells

3. Assign videos to analysts

1. Crowdsource videos from community

Urgency Map

Analyst

Server2. Upload relevant videos

Coverage Map Space Decomposition

Video contentVideo metadata

f1f2

f3

Video locationWork cellA video has high VA if

Enclosing work cell is urgentVideo coverage is large

3-seconds video

ResultsSynthetic Datasets [Ay et. Al. SIGSPATIAL’2010]

Visual Awareness Maximization (Uniform Dataset)

Statistics ValueSpatial distributions Uniform, Gaussian and Zipfian#videos 1000 with varying lengthRegion size 10 × 10 square km in Los Angeles

Video Level

16 25 36 49 640

0.51

1.52

2.53

3.5GridQuadtree

Analyst count

Visu

al A

war

enes

s

Kd-tree increases visual awareness by 300%

16 25 36 49 640

20

40

60

80

100

120

GridQuadtreeKd-tree

Analyst count

Visu

al A

war

enes

s

VA is 10 times higher in comparison to video-level

Frame Level