1. Recalibration Problem Virtual Sensors (VS’s) Detect Events E.g., Coffee Level Rescue Kitchen, West-East Traffic light Next to Calit(2). VS-Output Coffee State: 11:43 Half-Full 11:44 Half-Full 11:45 Half-Full 11:46 Half-Full. Traffic Light State: 11:43 Red 11:44 Red 11:45 Red 11:46 Red.

2. Recalibration Problem (Cont.) Perturbations in Physical Sensors Can Cause Virtual Sensors to Become Un-calibrated E.g, camera can move its FOV. VS-Output Coffee State: 11:43 Empty 11:44 Empty 11:45 Empty 11:46 Empty. Traffic Light State: 11:43 Orange 11:44 Orange 11:45 Orange 11:46 Orange.

3. Recalibration Problem (Cont.) Recalibration refers to detection of such perturbations & adjustment of parameters used by VS’s to detect correctly.

4. The Recalibration Problem - Approach Observation: Events = change in value of attributes. Approach: Model the state changes as an automata. Use the automata to capture semantics. AvgStdavg-2stdavg+2std O->R3.250.9571.335.16 G->O14.756.1852.3827.11 R->G154.7615.4724.52

5. Exploiting Semantics for Re-Calibration Detection: Deviation from the Learnt Model Correction & Calibration: A Search Over Parameter Space to Find Semantic Model DEMO: Finding the Correct FOV Based on System Semantics: Coffee Machine & Traffic Light

6. Exploiting Semantics for Scheduling of Resources Let: C be a Set of n Cameras. Plan (t) be a {0,1} vector (Indicating Which Cameras to Probe) Benefit(Plan (t) ) be the expected benefit from executing this plan. Cost(Plan (t) ) be the cost associated with that plan. Find a plan, such that: Benefit(Plan (t) ) is maximized. Cost(Plan (t) ) is minimized. Under the constraint: Σb i ≤ k For example, probe a set of N entry/exit sensors: To find as many distinct people entering. When establishing a new connection to a sensor is 5 time more expensive than continuing to probe the same sensor. Only k connections can be established at the same time.

7. Semantics Exploited Apriority: What are the cameras that I should probe, when nothing else is known. Self Correlation: If we’ve seen someone walking in camera A, how likely is it to see more motion in A. Cross Correlation: If this is the state of the system we are aware of: We’ve seen motion in camera B 5 seconds ago. And motion in camera C 2 seconds ago. How likely is it to see motion in camera A.

8. Motion Semantics Semantics Used: Motion Other Types of Semantics Trajectory Identity

9. Scheduling Using Semantics Naïve: RR Exploiting Sensor Semantics: 1. Determine Probability of Motion 2. Use the Motion Estimates to Find the Best Plan

10. Results Almost 90% improvement compared to RR!!

11. Different Benefit Functions The semantics based algorithm proves to perform the best...But it can do better if we change the plan based on the cost/benefit functions.