Czech researchers help crowd surveillance drones detect trouble using neural networks

drone surveillance crowds

Researchers in the Czech Republic have developed a way to make uncrewed aerial vehicles (UAV) more effective in monitoring large groups of people. To achieve that, scientists are making drones smarter by linking them to neural networks that detect when crowds under surveillance show signs indicating trouble.

Czech police drones in crowd surveillance taught to identify anomalies

The innovation comes from a partnership between the Brno University of Technology and Czech police, who have increasingly relied on drones in the surveillance of crowds. According to a report by Radio Prague International, the problem with UAVs in that role is their limited function as simple conduits of video to watching officials. By the time those images reveal fights, surges, panic, or other threats to safety, the danger they embody is already established and spreading. 

According to David Bažout, a graduate of Brno University’s Information and Technology school, “no one is able to assess what is normal behavior and what not.” As a result, police are left reacting to and trying to catch up with already unfolding emergencies. To remedy that, Bažout developed a neural network to analyze video from UAV that – much like a human brain ­– detects sudden changes or anomalies indicative of brewing trouble.

General footage from drones on surveillance missions of crowds is first divided into smaller groups, then reassembled into data representing a broad picture of normal activity. Accumulation of that recurring, regular behavior becomes the baseline from which any significant variation is distinguished. Those alterations are immediately identified and flagged to humans monitoring the situation by subjects of interest being highlighted in red. The system also features a self-learning element, meaning it – using drones as its eyes – can continually refine its appreciation of regular activity compared to sudden changes indicating potential problems.

Self-learning networks make drones smarter

Bažout says initial development used simple situations into which abrupt changes were introduced. One involved a group of children running in a counterclockwise circle, then stopping to reverse direction. Another filmed a soccer match whose typical action was suddenly punctuated by a few players lying down on the ground. In both cases, the neural network pegged the anomalous changes and flagged them in red.

Now Czech police are training with the drone-based system in their surveillance of crowds. But because individual human activity within groups can be unpredictable or unusual without necessarily being dangerous, initial trials provoked a great deal of otherwise unremarkable red. In response, Bažout and his research partners refined the platform with a gauge allowing cops to increase or decrease the sensitivity of what sets off the alarm signal. 

The main use of the system will obviously be in monitoring crowds, and allowing cops or first responders to identify sudden changes indicative of trouble and start responding before dangerous situations have time to evolve. Bažout says the platform may also be deployed in chasing fleeing criminals who duck into crowds to avoid detection and capture. The fast, dodging movements and evasive behavior of crooks on the lam, Bažout says, would allow the network to spot them as outliers.

The true test for the smart drone application would come monitoring US events like Burning Man, where virtually every movement and individual is both unexpected and strange.

FTC: DroneDJ is reader supported, we may earn income on affiliate links


Subscribe to DroneDJ on YouTube for exclusive videos

Show More Comments