A Highly Efficient Computer Vision Approach for Tennis Ball Retrieval with Limited Computational Power

Abstract

Collecting scattered tennis balls after training sessions is often a tedious task for already exhausted players. In this paper, we propose and compare two computer vision approaches that can be implemented on a relatively affordable mobile robot to automate the task of tennis ball collection. After investigating the utility of Convolutional Neural Networks (CNN) for performing the visual processing required to recognize and track tennis balls, we show that such deep learning systems can be too computationally demanding for simple embedded systems. We therefore shift towards a more classical computer vision approach, using basic filtering techniques from OpenCV, to prioritize the efficiency and simplicity of the vision algorithm. Classical outline detection techniques, in parallel with color and noise filtering, are found to be sufficient for accurate tennis ball detection at real-time speed. The objective of this paper is to integrate efficient computer vision technology with mobile robots and a reliable control system in order to solve the problem of tennis ball collection. Upon careful experimentation and evaluation of the deep learning and the OpenCV approaches, we conclude that the OpenCV based algorithm provides competitive results to the deep learning model while drastically improving energy efficiency, cost savings, and performance speed for real-time detection under a range of lighting and distance scenarios.