基于YOLOv5算法的水位智能监测系统

Intelligent Water Level Monitoring System Based on YOLOv5 Algorithm

摘要: 近年来，图像识别技术快速发展，为了实现水位数据前端处理，本论文提出了将YOLOv5算法用于前端处理水位监测数据的方案。本方案与传统方案最大的不同之处在于数据处理方式，本方案直接将YOLOv5算法嵌入集成单片机中，在前端进行数据转换，最后将结果显示在终端，传统方案需要将数据传输回服务器中。论文实验数据的采集，考虑了不同测量地点和光照角度等因素，具有真实性。最后将YOLOv5算法的实验数据与YOLOv4、hog + svm和Faster RCNN算法进行对比，并以F1作为比较标准。结果显示：在验证集下，YOLOv5算法的mAP值和准确率等指标均表现优异，F1值高于Faster RCNN将近4%，高于HOG + SVM将近12%，在GPU和CPU环境下，单张图片检测时间分别为15 ms、125 ms，说明YOLOv5算法在各种复杂环境下的均能适应图像识别任务，具有高识别率和鲁棒性。

Abstract: In recent years, image recognition technology has developed rapidly, in order to achieve front-end processing of water level data, this paper mainly proposes a scheme of using YOLOv5 algorithm for front-end processing of water level monitoring data. The biggest difference between this scheme and traditional schemes lies in the data processing method. This scheme directly embeds the YOLOv5 algorithm into the integrated microcontroller, performs data conversion in the front-end, and finally displays the results in the terminal. Traditional schemes require data transmission back to the server. The collection of experimental data in the paper takes into account factors such as different measurement locations and lighting angles, and has authenticity. Finally, compare the ex-perimental data of YOLOv5 algorithm with YOLOv4, hog + SVM, and Faster RCNN algorithms, and use F1 as the comparison standard. The results show that under the validation set, the YOLOv5 algorithm performs excellently in terms of mAP value and accuracy, with an F1 value nearly 4% higher than Faster RCNN and nearly 12% higher than HOG + SVM. In GPU and CPU environments, the single image detection time is 15 ms and 125 ms, respectively. This indicates that the YOLOv5 algorithm can adapt to image recognition tasks in various complex environments, with high recognition rate and robustness.