A Lightweight Convolutional Network for Deep Learning-based Ball Manipulation by a Soccer Playing Robot

Automated detection, recognition, tracking and manipulation of objects by a mobile robot are valuable for entertainment, service and surveillance applications. Entertainment applications include soccer-playing robots while robots engaged in handling cutlery, medicines and food form part of the service segment. With the advent of deep learning, many conventional tasks can now be performed with higher accuracy. A convolutional neural-network-based hardware-software codesign for robotic object detection, tracking and following is presented. The proposed convolutional neural network is lightweight and is appropriate for realization on an embedded platform with constraints on hardware. An indigenously developed low-cost wheeled robot equipped with a vision sensor is described and a soccer-playing application based on deep learning is demonstrated with the robot. The object detection and following task is executed on the mobile robot (equipped with a Raspberry Pi) at a frame rate of 3.2 frames per second.