AI researchers are working on a project that will teach drones to analyze footage of crowds and root out violence. The project uses ScatterNet Hybrid Deep Learning to allow drones to pick humans out from aerial footage and look for violent acts. In order to do this, the drones use pose estimation, wherein they look at people in video footage and use a wireframe skeleton model to map out their joints and look for actions and poses that are similar to a database of violent acts that the AI was trained on. Naturally, this database will grow over time and is currently in a fairly rudimentary state.
The researchers on the project include Amarjot Singh from Cambridge University UK’s Department of Engineering, Devendra Patil from India’s National Institute of Technology, SM Omkar from the Indian Institute of Science. Bringing together knowledge of biology, AI-based image processing and machine learning, the three have created a system that’s able to identify staged examples of violent behavior with an 85% accuracy rating. The initial training data set includes action data for strangling, punching, kicking, shooting, and stabbing. This leaves a lot of potentially violent actions that could happen in a crowd on the table, such as ground grappling, throwing breakable objects or explosives, or swinging large objects like sticks and street signs. All of the data processing and AI magic happens in the cloud, which means that the system is easily scalable to include more drones and more processing power. This, in turn, will make it easier for the system to learn more about human violence, and to analyze more people at once and a wider range of footage simultaneously.
This technology is an advancement in the wider area of mass surveillance via drone, a subset of computer vision and machine learning that’s been gaining ground lately. Governments in some parts of the world, such as China, make use of a number of mass surveillance technologies to keep track of people and root out criminal behaviors on the streets. As to more specific applications, one can look to the US Military’s Project Maven for a sterling example of this sort of technology being used and enhanced.
