FYI.

This story is over 5 years old.

Tech

Machine Learning Will Save India's Cows from Bad Drivers

A road construction explosion in India is pitting heifers against highways.

In certain parts of the world, cow collisions are the stuff of legend. A car traveling at any significant speed up against a full-grown cow travelling at no speed is most often the end of the car, the cow, and maybe the human(s) inside of the car. That's up to 1,800 pounds of beef standing there with a center of gravity well-optimized to ensure that those pounds wind up where they can be most dangerous.

Advertisement

Cows on roads are a big problem in India, where rapid urbanization and industrialization has meant that new roads are increasingly being laid through rangeland. A 2015 study found that some 6 percent of accidents in India can be attributed to animals on the road.

Enter Indian computer scientists Sachin Sharma and Dharmesh Shah. The duo has crafted a cow identification algorithm that, if deployed, could reduce the incidence of cow-car collisions, they argue. It's described in the International Journal of Vehicle Autonomous Systems.

"Road crashes have been a major problem in India in recent times," Sharma and Shah write. "The occurrences have increased considerably owing to the influx of fourwheelers and two-wheelers on roads. The interior roads connecting villages and towns have been instrumental in multiple animal–vehicle collisions. Though numerous efforts have been in progress to reduce the number of collisions, lack of practical applications has impeded any major breakthrough in the scenario."

This shouldn't be all that mind-blowing for those who follow AI news. It's basically just a special case of object recognition—the cow case. So, an algorithm is shown a whole bunch of pictures, some with cows and some without. Eventually it produces a model or abstraction of the features that most make a cow a cow and not something else. Feed this model along with some new data (video frames taken from a dashboard camera) into some equations and we wind up with a prediction of cow or no-cow. This is more or less how computers detect faces or any other object.

The complete system incorporates both a detection system and a warning system. What makes it better than, say, a range sensor (which would catch any obstacle with no computation needed) is unclear. A warning that only flags cows and not other obstacles actually seems kind of dangerous, doesn't it?