Coopetitive Multi-Camera Surveillance using Model Predictive Control
Basic Idea Conceptual Motivation Key Contributions

Practical Implementation
Dual Camera setup

Practical Implementation
Triple Camera Setup
People Publications

Practical Implementation/Experimentation

b) Triple Camera Setup

We are looking at a triple camera setup to monitor any given walkways for "abandonement" activity. Our experimental setup is shown in [Figure 2]. In such an evironment, we want to continously monitor for any abandonement event, after which we want to obtain the following information:-

i) Three facial images of the person abandoning the object - Who is he?
ii) Three images of the object - What has he abandoned? and
iii) The trajectory covered by the person - What might be his intention?

The three cameras are allocated one of the three above mentioned subtasks based on their suitability. In this setup competetion is required so that the best suited camera is allocated as FaceCam i.e. "the Camera focusing on the Face". Cooperation is required so that the ObjectCam i.e. "Camera focusing on the object" can obtain correct location information from the camera which detects the event. The role allocation is dynamic and done automatically by the system.

We again compare our Coopetitive+MPC approach with other plausible approaches. We found that our proposed approach performed better than other approaches.

Approach Cam1
Role Allocation 3 Facial
3 Object Image Face capturing speed Demo Video
Only Cooperation Face Object Trajectory Incorrect No Yes - Video
Only Competition Trajectory Object Face Correct Yes No - Video
Coopetition+PID Trajectory Object Face Correct Yes Yes Slow Video
Coopetition+ MPC Trajectory Object Face Correct Yes Yes Faster Video

The "Only Cooperation" employs no measure of suitability and assumes all sensors to be equally capable. This results in unsuitable sensor being allocated as FaceCam. The "Only Competition" approach on the other hand employs no transfer of information between sensors hence ObjectCam can not obtain information about location of the object. Also, analysis of our data showed that use of MPC makes the capturing of three facial images faster as compared to PID (Proportional Integral Differential) control.