This was my bachelor thesis at Qazvin Azad University, and a direct continuation of the problem I'd been working on at the NAO Biped Lab for years: detecting objects reliably in a soccer environment, on hardware that couldn't afford to be slow.
The problem
Detecting a soccer ball sounds easy until you're doing it on a NAO robot during a RoboCup match. Lighting varies significantly across venues. The ball moves. Other robots occlude it. And you have a hard real-time budget — the vision pipeline needs to complete fast enough for the locomotion system to react.
The earlier approach was color-based: segment the image using color thresholds, apply heuristics. It worked but was brittle, especially under unusual lighting. The thesis question was whether CNNs could handle this more robustly, at competitive speed.
The approach
We trained convolutional networks for detection and evaluated them under the constraints that mattered: accuracy across different lighting conditions, inference time on the robot's hardware, robustness to partial occlusion. CNNs handled the lighting variation much better than the color-based approach. Getting the model small enough to run at the required frame rate required careful architecture choices, but it was achievable.