You Only Look Once (YOLO) – Based Human Detection for a Scuttle Bot

Abstract

One of the most important applications of a mobile robotics system is real-time human detection especially where safety and human-robot interaction are very crucial. This work involves the problems of applying the deep learning-based detection system to a limited resource robotic platform with the You Only Look Once (YOLO) algorithm. It was also deployed on the SCUTTLE mobile robot, e.g., the single-board computer (Raspberry Pi 4) uses a USB camera and  and ROS2 middleware is used. Since a publicly available construction safety dataset was used to train the system, it could be taught to identify human presence and personal protective equipment (PPE), like hardhats and safety vests. Data gathered and performance assessment were done with the aid of pre-recorded video instead of live cameras since the system is still in its prototyping stage. The first tests were performed on desktop PC with PyCharm after which similarities and differences between platforms were observed with Raspberry Pi 4. Although both platforms performed well in terms of detection accuracy (over 85 percent in indoor, controlled settings), the Raspberry Pi was limited by frame rate (1-2 FPS) and inference latency, whereas only low-frequency monitoring could be performed without optimization. The work shows the possibility of fusion object detection real-time on embedded mobile robot and gives an idea about how to overcome the balance limitations of accuracy and performance of such edge-based AI solution in the robotics field.