Crowdsourcing the Acquisi3on and Analysis of Mobile Videos for Disaster Response

Transcription

1 IEEE Big Data 2015, October 31, 2015 Crowdsourcing the Acquisi3on and Analysis of Mobile Videos for Disaster Response Presented by Hien To Dr. Seon Ho Kim Integrated Media Systems Center University of Southern California Prof. Cyrus Shahabi 1

2 GeoQ by Na3onal Geospa3al- Intelligence Agency (NGA) GeoQ Each Analysts project job is a evaluate contains crowdsourcing contains a mobile set a of set plamorm videos work of jobs cells, within to which analyze their are assigned disaster data work to analysts cells 2

3 Mo3va3on ü Obtain real- 3me informa3on to enhance situa*onal awareness during and aser disasters ü Live streaming of mobile video, e.g., YouTube, Periscope ü Priori*ze data collec*on with limited transmission ü Analyze collected data in 3mely manner 3

4 Disasters Assess damage across large geographical area Hai3 Earthquake Cable Cuts in Vietnam Disrupted communica3on networks or data overload! Hai3 earthquake, 85% Hai3ans could s3ll access to their mobile phones but 70% cell towers were destroyed 4

5 Crowdsourcing Disaster Response Early disaster data collec3on focus on VGI Disaster surveillance and response system Processing and integra3on of disaster data by leveraging Linked Open Data [Goodchild et al. 2010] [Chu et al. 2012] [Ortmann et al. 2011] Combine human and machine intelligence [Schulz et al. 2012] Infrastructure damage caused by Earthquake [Kobayashi 2014] Resilient communica3on technologies, e.g., drones Our study: [Nemoto and Hamaguchi 2014] ü Focuses on both data collec3on and data analysis ü Collects huge amount of videos concerning disasters ü Improves resilience and responsiveness of computer systems 5

6 Crisis Response Efforts Ushahidi AI for DR USGS ShakeMap Did You Feel It? USC MediaQ NGA GeoQ Planetary Response Network Highlight GeoQ and MediaQ 6

7 p Data Collec*on MediaQ [1]: mobile crowdsourcing plamorm to collect videos/images N α R MediaQ and GeoQ Field Of View (FOV) FOV Metadata: d p: camera locamon d: camera direcmon α: viewable angle R: viewable distance t: Mmestamp ü MediaQ provides visual data to GeoQ. Data Analysis GeoQ [2]: crowdsourcing plamorm to analyze disaster data Work cells are assigned to analysts GeoQ s Work Cells used in Nepal Earthquake 2015 ü MediaQ workers are on- site people; GeoQ workers are off- site analysts. [1] By USC Integrated Media System Center [2] By US NaMonal GeospaMal- Intelligence Agency (NGA) 7

8 Proposed Techniques Collect data fast because 3me is cri3cal Analyze the collected data effec3vely Coverage Map Video metadata ü Upload metadata (FOV) first then iden3fy videos with high visual awareness to be uploaded Video content ü Par33on space based on video loca3ons 8

9 Proposed Framework Coverage Map Space Decomposi3on Server 2. Upload relevant videos 3. Assign videos to analysts 1. Crowdsource videos from community Analyst 0 4. Evaluate work cells Video metadata Video content Urgency Map 9

10 1. Acquisi3on of Video Metadata and Content ü Hold video in the device and have server ping the device when it wants the owner to upload the video ü Real- 3me acquisi3on of metadata via Internet, SMS, WLAN ü Size of metadata is thousands 3mes smaller than content ü Real- 3me analysis on the uploaded metadata, e.g., visualiza3on ü Support data governance, preserve privacy and strict access control [Dey and Chinchwadkar, ICDE 2015] 10

11 2. Visual Awareness Maximiza3on Visual Awareness (VA) of a Video VA(v) =Urgency(w) area(v) area(w) f 1 f 2 f 3 A video has high VA if ü Enclosing work cell is urgent ü Aggregate video coverage is large Video location Work cell 3- seconds video within a work cell Selects videos that maximize total VA without exceeding budget (i.e., constrained bandwidth) This problem is 0-1 Knapsack. 11

12 3. Assign Videos to Analysts GeoQ uses of grid- based par33on ü Some analysts may be overloaded L Space par33oning algorithm based on Quadtree and Kd- tree ü Each analyst handles equal number of videos J ü Terminates when #workcells > #analysts- 3 Kd- tree (Gaussian dist.) Quadtree (Gaussian dist.) 12

13 Minimum Redundant Coverage Upload key frames and metadata to server and minimize redundant coverage f 1 f 3 f 2 f 4 FOV overlap Select frames that maximizes total Visual Awareness without exceeding budget This problem is Maximum Coverage Problem. 13

14 Simula3on Results 1. Par33oning techniques: Grid, Quadtree, Kd- tree 2. Visual awareness at video level vs. frame level Synthe3c Datasets [Ay et. Al. SIGSPATIAL 2010] Sta*s*cs Value Spa3al distribu3ons Uniform, Gaussian and Zipfian #videos 1000 with varying length Region size square km in Los Angeles area Unit grid cells FOV # per second 1 Viewable angle 60 degrees Visible distance 200 meters 14

15 Visual Awareness Maximiza3on at Video Level Varying the number of analysts Variance of #videos per analyst Visual Awareness Grid Quadtree Kd-tree Analyst count Uniform Dataset Kd- tree increases visual awareness by up to 300% Variance (log scale) Grid Uniform Gaussian Zipfian Quadtree Kd-tree Gaussian Dataset Kd- tree produces similar #videos per work cell 15

16 Visual Awareness Frame Level Analyst count Grid Quadtree Kd-tree Uniform Dataset Obtained visual awareness is an order of magnitude higher in comparison to video- level opt Run3me Measurements Construction time (ms) O(n) 5.2 O(log An) O(n logn) Grid Quadtree Kd-tree Uniform Dataset Construc3on 3mes are small and the differences are insignificant 16

17 Conclusion ü Proposed crowdsourcing framework for collec3on and analysis of video data under disaster situa3ons ü Introduced a mechanism to enable 3me- sensi3ve acquisi3on and analysis of spa3al metadata ü Developed an analy3cal model to quan3fy the visual awareness of a par3cular video or frame ü Introduced two variants visual awareness maximiza3on problem: uploading the en3re videos and individual frames ü Introduced more effec3ve spa3al decomposi3on to assign videos to analysts 17

18 Q/A Hien To University of Southern California 18

19 MediaQ Dataset Visual Awareness Grid Quadtree Kd-tree Analyst count 19

20 Loca3on/Region Entropy Loca3on entropy (LE) Measures the diversity of unique visitors of a locamon High if many users were observed at the loca3on with equal propor3on l 1 u u LE 0.35 l = u LE( l) P( u)log P( u) U l where P ( u) = l l O Freq ( l) = O l UserCount ( l) =, l u, l O l O u l : total number of visits : number of visits by worker u Cranshaw, J., et al., (2010). Bridging the gap between physical locamons and online social networks. UBICOMP,

21 Work Cell Urgency Compute urgency value of a work cell Urgency(w) = Priority(w)α + RegionEntropy(w.r)β The priority of an area, e.g., nuclear plant areas are more important than residence areas. Region Entropy geosocial factor of a region. RE is high if the region are visited by many people many 3mes, e.g., school, hospital. Nuclear plant USC Campus 21

22 Future Work 1. Use user- generated videos from MediaQ 2. Use near- real- 3me ShakeMap from USGS 22

23 Data Analysis Work Cell Analysis ü Analysts measure the severity of their assigned work cells and set urgency values to them ü Urgency of the work cells can be automa3cally calculated 23