Researchers Develop Surveillance System That Can Watch & Predict

hypnosec writes "Carnegie Mellon university researchers have developed a surveillance system that can not only recognize human activities but can also predict what might happen next. Scientists, through the Army-funded research dubbed Mind's Eye, have created intelligent software that recognizes human activities in video and can predict what might just happen next; sounding an alarm if it detects anomalous behavior. "

Comments

Next up....

[CimmerianX]

Thought Police Alpha Version .501 right here. Arrest him!!!! Our system assumed he would shoot somebody.

Re:Next up....

[betterunixthanunix]

I suspect this system will see more subtle use. I suspect that this system will be deployed within corporations, and used to detect employees who are not satisfied with their treatment. Those employees will either be fired or promoted (to divide them from their peers and prevent them from organizing people). The purpose of a system like this is to enforce the social order, to prevent change, and to ensure those who are in power will remain in power.

Re:Next up....

[Jeremiah Cornelius]

You're Eight Times More Likely to be Killed by a Police Officer than a Terrorist
http://www.cato-at-liberty.org/youre-eight-times-more-likely-to-be-killed-by-a-police-officer-than-a-terrorist/print/

Will it be as intelligent as Siri?

[elrous0]

I said "The corner of VINE" not "PINE," you dumb bitch!

Here we go.

[vikingpower]

In my native Europe as much as in your great & free US: More surveillance state. More police state. More security craze. Where is this going to stop ? When are ordinary, yet intelligent people going to refuse to live in and contribute to such a state ?

It's vaporware

[Animats]

No, they haven't "developed a surveillance system". The paper is two psychologists blithering about the potential architecture of one. It reminds me of the awful papers that came from the "expert systems" community in the 1980s. There's been some progress; it mentions Bayesian statistics. But it's fundamentally an approach based on parsing visual data into something that looks like predicate calculus and grinding on that. There's a long history of that not working.

It's an idea in the right direction, though. A key component of intelligence is prediction. Knowing what is likely to happen is a basic component of common sense, an area in which AI systems have historically been weak. With prediction comes the ability to ask "what if" questions, essential to deciding what to do next without doing something stupid.

There's been real progress in that area, but not from the expert systems people. Adobe Photoshop's content-aware fill is an example of a successful system which has a form of "common sense" - it fills in plausible-looking areas to replace sections deleted from photos. Related technologies exist for videos, and are used for motion compression and 2D to fake 3D conversion. Systems which look at video and guess "what happens next" may be the next step.

fairly predicatable

[andrew_d_allen]

With most surveillance footage it's pretty easy to spot what's going to happen next: the customer will pay for their items, receive change, and walk out of the store. Unless you're watching it on the internet. Then, a car will drive into the storefront or a botched hold-up will occur.

Needs mass spook-spamming

[heretic108]

I'm thinking along the lines of the emacs "spook" function, amongst other things. You just need enough a large enough group of participants working together.

The system can be trained in weird ways. For instance, if enough people in enough places scratch their noses with their left hands, then break out in a mock fight, the system will learn to sound the alarm every time someone scratches their nose with their left hand.

Or, for something more socially useful - have people pull out a cellphone, talk for a few seconds, then pull out a mock gun and pretend to mug others. Then, the system will freak out every time some annoying jerk pulls out a cellphone in public. Along that same theme, train the system to send in the troops whenever someone adjusts their underwear in public, or picks their nose, or farts loudly...

Bogus Research...FTA

[globaljustin]

I looked through the full text of the research (http://stids.c4i.gmu.edu/papers/STIDSPapers/STIDS2012_T02_OltramariLebiere_CognitiveGroundedSystem.pdf)

It is bogus. Wouldn't get published. They say the system *predicts behavior* using a systematic behavior ontology. When it describes their theory of the ontology it lists three factors in the system.

"Causal Selectivity" #3 is the one that links **cause and effect** its the part of the equation where your action (reaching in pocket) is either interpreted as something threatening (trigger bomb) or non-threatening (scratch balls discretely in public).

Guess what...all they do is say "Will be addressed in further research"...!!!

The whole basis for their claim...'prediction' is explicitly not part of this research. They do not even address the link of one behavior to another, yet it is the whole premise of their claim!

From page two (emphasis added)

Ontology pattern matching - comparing events on the basis of the similarity between their respective pattern components: e.g., a person’s burying an object and a person’s digging a hole are similar because they both include some basic body movements as well as the act of removing the soil;

Conceptual packaging - eliciting the conceptual structure of actions in a scene through the identication of the roles played by the detected objects and trajectories: e.g. if you watch McCutchen hitting an homerun, the Pittsburgh Pirates’ player number 22 is the ‘agent’, the ball is the patient, the baseball bat is the ‘instrument’, toward the tribune is the ‘direction’, etc.).

Causal selectivity: attentional mechanisms drive the visual system in picking the causal aspects of a scene, i.e. selecting the most distinctive actions and discarding collateral or accidental events (e.g., in the above mentioned homerun scenario, focusing on the movements of the rst baseman is likely to be superuous). In the next section we describe how the Cognitive Engine realizes the rst two functionalites by means of combining the architectural features of ACT-R with ontological knowledge, while **Causal selectivity will be addressed in future work.**