Player Detection

For player detection, we use a single-stage object detection CNN based on the SSD meta-architecture, which is capable of detecting robots in an end-to-end fashion. With respect to the widely popular MobileNetV2-SSD, our model is much faster - up to 20-30 FPS on the embedded CPU of the NAO V6. This was achieved by using a custom light-weight backbone, resizing the input size to 120x90, and using only the Y-channel.

The model was trained on the JETNet dataset using Tensorflow Keras. It is deployed using Tensorflow Lite.