基于YOLOv5和树莓派的乒乓球拾球移动小车系统设计

Ping-pong Ball Picking Mobile Cart System Design Based on YOLOv5 and Raspberry Pi

为了高效准确地提取乒乓球图像特征,对在复杂背景下的两种乒乓球进行精确识别。以网络上乒乓球图片为样本,利用单阶段目标检测模型YOLOv5对乒乓球数据进行训练,将训练好的模型部署于Windows平台,并且利用motion通过树莓派连接外接摄像头进行拍摄并传输到Windows端,Windows端将识别结果(乒乓球的坐标)通过传输控制协议(Transmiss ion Control Protocol,TCP)传送给树莓派,最后树莓派控制小车将乒乓球捡入收集篮中。结果表明:本研究训练的模型平均精准度高,AP50指标和AP@50:5:95指标分别为97.6%和67.0%。Windows和树莓派端程序稳定可靠,在体育馆等复杂环境下,可以快速、准确地识别乒乓球并且定位乒乓球的位置,小车可以精准地捡球并将其放入收集篮中。本研究训练的模型具有稳健性强、实时性好、精准度高等优点,系统可以长时间并且稳定地拾取乒乓球。

To efficiently and accurately extract table tennis image features for accurate recognition of two types of table tennis balls in complex backgrounds. Taking the table tennis pictures on the network as samples, the single-stage target detection model YOLOv5 is used to train the table tennis data. The trained model is deployed on the Windows platform, and motion is used to connect the external camera through the raspberry pie for shooting and transmission to the Windows end. The Windows end transmits the recognition results(table tennis coordinates) to the raspberry pie through the Transmission Control Protocol(TCP), Finally, the Raspberry Pi control car picks up the table tennis ball into the collection basket. The results show that the average accuracy of the model trained in this study is 97.6% and 67.0% for AP50 and AP@50:5:95 respectively, and that the Windows and Raspberry Pi programs are stable and reliable in identifying and locating ping pong balls quickly and accurately in complex environments such as gymnasiums. The model trained in this study has the advantages of robustness gun, good real-time performance and high accuracy,and the system can pick up ping pong balls for a long time and in a stable manner.