Structural Pruning for Real-Time Multi-object Detection on NAO Robots

Abstract

In this paper, we propose a real-time multi-class detection system for the NAO V6 robot in the context of RoboCup SPL (Standard Platform League) using state-of-the-art structural pruning techniques on neural networks derived from YOLOv7-tiny. Our approach combines structural pruning and fine-tuning, to obtain a pruned network that maintains high accuracy while reducing the number of parameters and the computational complexity of the network. The system is capable of detecting various objects, including the ball, goalposts, and other robots, using the cameras of the robot. The goal has been to guarantee high speed and accuracy trade-offs suitable for the limited computational resources of the NAO robot. Moreover, we demonstrate that our system can run in real-time on the NAO robot with a frame rate of 32 frames per second on $224\times 224$ input images, which is sufficient for soccer competitions. Our results show that our pruned networks achieve comparable accuracy to the original network while significantly reducing the computational complexity and memory requirements. We release our annotated dataset, which consists of over 4000 images of various objects in the RoboCup SPL soccer field.