Academia Up With Industry Improves Intelligent Surveillance Systems

Contribute to improve surveillance systems is the goal of a research project carried out by Francisco Melo, assistant professor at Instituto Superior Técnico of the Universidade Técnica de Lisboa and senior researcher at INESC-ID, and Manuela Veloso, researcher at Carnegie Mellon University, with the Portuguese company ObservIT.

The MAIS+S - Multiagent System for Intelligent Surveillance project aims to use current research in multi-agent planning and apply it to a specific real-world scenario - the creation of intelligent surveillance systems.

Francisco Melo considers that the work with ObservIT is being very proactive, “not only our research has the potential of greatly improving the performance of their surveillance networks in several aspects, but also the discussions that often take place in our frequent meetings have led to many interesting ideas that can be explored in their systems or in future collaborations.” Carried out in the scope of the Carnegie Mellon Portugal Program, funded by the Portuguese Foundation for Science and Technology, this research project is now getting to its third year.

 Francisco S. MeloWe spoke with Francisco Melo about the real-world testbed that it is being shaped at ISR, the achievements of the project until know, and about the connection with ObservIT.

Carnegie Mellon Portugal Program (CMU Portugal): Why is the MAIS+S project important?
Francisco Melo (FM):
The overall goal of the MAIS+S project is to use current research in multi-agent planning and apply it to a specific real-world scenario - the creation of intelligent surveillance systems. There are two important motivations behind this goal:
- From a research perspective, the multi-agent planning approach adopted in the project (known as decision-theoretic planning) has been the focus of much recent research. Decision-theoretic planning in multi-agent settings is a very hard problem, and much of the recent work has focused on particular instances of this general problem that exhibit some structure that can be used to simplify the problem. Surveillance networks have natural structure that can be effectively exploited. Moreover, decision-theoretic planning research has placed a great emphasis on theoretical results, but has seldom been applied to real-world problems. MAIS+S provides a real-world testbed for the theoretical research conducted within the project, enabling a strong empirical assessment of the validity of common assumptions in this area of research.
- From a technology-transfer perspective, MAIS+S work has been developed in close collaboration with the surveillance company ObservIT. Our collaboration has been very fruitful, and our research will contribute to improve surveillance systems in several aspects.

CMU Portugal: Could you give some examples on how the results of this research project can be applied?
FM:
I can give three examples. The first example is the robustness to failures in the network. Given the decentralized nature of the algorithms developed within MAIS+S, the agents in the network - robots, security personnel, control central - are able to coordinately respond to detected events, even if a failure takes place in parts of the network. The second example is the possibility to extend surveillance capabilities. The use of robotic agents integrated in the network allows the network to survey spaces not covered by the fixed cameras. For example, if a person of interest is detected, robots can be dispatched to ensure that such person is never out of the field-of-view of the network. The third example is related to smart event handling. The information displayed to the personnel in the control central is controlled by intelligent algorithms that are able to prioritize the different events detected by the network, allowing for more efficient management of events by the human controllers. Also, the robots in the field, are able to prioritize the different events detected and autonomously respond to the most urgent events.

CMU Portugal: More than one year after its start, what were the main achievements of the research team?
FM:
The first year of the project was focused on research. The project produced a number of significant results, both in terms of decentralized planning and in terms of vision. Namely,
- The state-of-the-art optimal Dec-POMDP solver was developed in the context of the project. A Dec-POMDP is a complex model of a multi-agent system used in decision-theoretic planning, and investigated extensively in MAIS+S research.
- We developed a new algorithm that, for special classes of Dec-POMDPs (such as those considered in the project), is able to solve problems orders of magnitude larger than other existing methods.
- We developed distributed learning algorithms that enable an agent in a network of agents to adapt to the actions of other agents.
- We developed a new vision algorithm that allows a camera network to track a person across multiple cameras.
During this second year, although maintaining the strong research profile, we have been working close to ObservIT to define test scenarios where the research could be applied. We have acquired and deployed the necessary hardware, and the first initial scenario should be in place within the next month. The last year will be focused on deploying the remaining (more complex) scenarios.

CMU Portugal: How do you comment on the work carried out with Observit?
FM:
The work with ObservIT has been fundamental to ground the research within MAIS+S. It is common for researchers to rely on assumptions that, although enabling powerful theoretical results, are often challenging to deploy in real world scenarios. Our frequent meetings with ObservIT, where we always present the most recent results in terms of research, and discuss the impact on surveillance networks, has helped our research to remain grounded on real world systems, contributing to make our results more solid. Also, from ObservIT's perspective, I believe that MAIS+S has contributed to set ObservIT apart from their competition in terms of surveillance systems. Not only our research has the potential of greatly improving the performance of their surveillance networks in several aspects, but also the discussions that often take place in our frequent meetings have led to many interesting ideas that can be explored in their systems or in future collaborations.

November, 2012