Cognitive Pilot experts have come up with an approach that allows a neural network to automatically select diverse and representative data from a video stream, the company said.
The new solution will allow developers to save tens of man-years in the development of autopilots, as well as greatly simplify the development process in this area, the company notes.
The authors of the development solved an urgent problem in AI for autopilots – the selection of various, representative data from a video stream. This task can be very time consuming in applications such as autopilot.
The authors have developed an AI that recognizes objects of a road, field or other scene with industrial accuracy: this is necessary to ensure safety in all weather conditions and time of day.
The mechanism developed by the Cognitive Pilot specialists also allows filtering out “garbage” data from the video stream that does not affect the learning process in any way – such as moving the combine from one field to another with the header raised, or moments when it stands still or passes a section of the route with disdain. small changes in the field scene.
“If earlier it could take years to process a video stream when training neural networks and creating datasets, today we press a button and we get the result,” concludes Minkin.