1. http://openisdm.iis.sinica.edu.tw Jane W. S. Liu Institute of Information Science, Academia Sinica Fusion of Human Sensor Data and Physical Sensor Data Edward T. H. Chu Computer Science Department, National Yun Tech University P. H. Tsai Inst. of Manufacturing Info. Sys. National Cheng-Kung University

2. Disaster Surveillance System  Functionalities include to  Collect real-time data  Acquire situation awareness  In-situ physical sensors include  Surveillance cameras  Wireless sensors  Inadequate coverage can be a problem

3. Cases for Crowdsourcing Full sensor coverage too costly In-situ sensors damaged

4. Social & Participatory sensing  Increasingly broader everyday applications, e.g., Mapping air quality, ambient noise, bus comfort, snow drift, etc.  Increasing more useful tool for disseminating information during disaster response e.g., use of twitter, Ushahidi and Sahana Crowd-Driven Crowdsourcing

5. Crowd- vs System-Driven Crowdsourcing  Shortcomings of crowd-driven crowdsourcing  Some locations may be visited by more participants than needed; others by too few, prolonging time required to explore the disaster area;  May expose participants to danger.  Elements of system-driven crowdsourcing  Use motivated participants known to the system;  Assign selected participants to locations to maximize coverage and other figures of merits;  Plan an exploration tour for each selected participant;  Manage each CDC (crowdsourcing data collection) process actively.

6. CROSS Command Center Locations & Observations (CROwdsourcing Support system for disaster Surveillance) Participants Assignments & Directions

7. Major Components of CROSS  Broadcast manager  Post CFP with participant requirements  Communicate with selected participant  Path planning manager  Compute a route for each selected participant  Redirect participants when needed  Crowdsourced map manager  Integrate the reports from participants and display it on a map  Update boundary of threatened area  Detection/estimation tools for determining quality of sensor coverage

8. Elements of Sensor System (Existing) physical sensors Missing sensors: a location where a physical sensor should be, but there is none Virtual sensor: A set of human sensors reporting from a missing sensor. Human sensor: a participant who sends his/her observations (human sensor data) during a CDC process

9. Characteristics of Sensors  Physical sensors are functionally identical:  They observe and send the same type(s) of data: e.g., Tar-ball-present, Water-over-the- curb, {wind direction, wind speed, humidity}  Their sample values = true values + noise  Noises of different sensors are statistically independent, identically distributed  Virtual sensors are similar to physical sensors:  They contribute the same type(s) of data  Additive noises in samples from human sensors are statistically independent, identically distributed

10. Alternative Goals of Fusion  Reliable detection of conditions that require timely response, e.g.,  Tar-ball-present → Launch clear up  Wind-in-direction-of-wildfire → Close park  Water-level > 10 cm → Close road (Binary hypothesis testing) versus multiple hypotheses testing  Accurate estimation of decision support parameters Future work

11. Binary Hypothesis Testing L(y)  Observables for K sources : random vector Y = (Y 1, Y 2 …Y K )  Input: a sample y = (y 1, y 2 …y K ) of Y  Output: Select hypothesis H 1 (H 1 is true) or H 0 (H 1 is false)  Assumptions: a prior probabilities are unknown; conditional probabilities Pr [ y | H 1 ] and Pr [y | H 0 ] are known.  Test statistics: likelihood ratio: L(y) = Pr [y | H 1 ] / Pr [y | H 0 ] Pr [L(y) | H 0 ] Pr [L(y) | H 1 ] Detection threshold

12. Neyman-Pearson (N-P) Criterion p Pr [L(y)  η* | H 0 ] + (1 - p) Pr [L(y)  η | H 0 ] = α N-P probabilistic rule: select  H 1 if L(y)  η;  H 0 if L(y) < η*;  H 1 with probability p, H 0 otherwise if L(y) = η* Given false alarm threshold f ≤ α, maximize detection probability d Pr [L(y)  z | H 0 ] = α L(y) Pr [L(y) | H 0 ] Pr [L(y) | H 1 ] η η*η* z

13. Neyman-Pearson (N-P) Criterion α ≡ accept false alarm probability η and η* = two adjacent values in the set Λ of all values of L(y) and Pr [L(y)  η | H 0 ] ≤ α, Pr [L(y)  η* | H 0 ] > α L(y) Pr [L(y) | H 0 ] Pr [L(y) | H 1 ] η η*η* Probability p is such that p Pr [L(y)  η* | H 0 ] + (1 - p) Pr [L(y)  η | H 0 ] = α N-P probabilistic rule: select H 1 if L(y)  η; H 0 if L(y) < η*; H 1 with probability p, H 0 otherwise if L(y)  η N-P test maximizes detection probability d for a given false alarm threshold f ≤ α (α, β) – coverage is achieved if d  β

14. Decision Versus Value Fusion  Decision fusion  Each sensor makes a local decision based on its own observation  Fusion center makes an overall decision based on decisions of individual sensor  Use for sensors with high data volume & have large processing capability  Value fusion  Decision is made on the basis of observed values of all sensors  Value fusion for all sensor is better

15. Centralized Decision Fusion: to achieve (α, β) - coverage 1 do decision fusion for physical sensors with overall false alarm threshold F = α; compute U and D; 2. if D  β, go to take action according to overall decision U; 3. broadcast Call-For-Participation; wait for responses; 4. from responded human sensors, select participants and allocate them to virtual sensors; 5. wait for a sufficient number of samples from human sensors assigned to each virtual sensor; 6. do value fusion for virtual sensors; for each, get local decision u i, false alarm rate f i and detection probability d i ; 7. do decision fusion for all sensors with overall false alarm threshold F = α to get overall decision U; 8. if D is less than β, send updated instruction to human sensors; goto step 5; 9. terminate the current CDC process; go to take action according to decision U;

16. On Binary Hypothesis Testing Based on Binary Sample Values Given inputs u i, f i, and d i, i = 1, 2, … K, Use N-P test to get U, with F and D. Are F ≤ min ( f i ) and D  max ( u i ) ? Answers:  Yes, if f i ≤ α for all i  If f i = α and d i  β for 3 or more inputs, then F = α, D  β Thomopoulos, S. C. A.,et al,, “Optimal decision fusion in multiple sensor systems,” IEEE Trans on Aerospace and Electronic Systems,,September 1987 More questions:  If f i = f > α, how many inputs are required to get (α, β) – coverage?  Can the same be said for arbitrary valued samples?

17. Binary Hypothesis Testing Multiple Hypothesis Testing Fusion Data from Unknown Human Sensors Building a Prototype Fusion Center Statistical Estimation Based on Symbiotic Sensor Data

18. Thank You!