Model Optimization and Integration to Jetson Nano
In tennis during a workout on the court is a lot of gaming balls. Usually, their number reach 200 objects and all of them must be removed manually. We were approached by the manufacturer of robots for assembling tennis balls during and after training in order to improve recognition for collecting balls on the court.
Since the robot functions with the Jetson Nano microcomputer, we needed to create an easy and fast solution. We used the compact version of the neural network Tiny-YOLO 3 to detect objects and the PyTorch YOLOv3 software. The neural network is trained on a set of 700 tagged pictures.
The neural network operates at a speed of 25 FPS and detects tennis balls with an accuracy of 98%.