复杂场景下无水尺水位的影像水位反演智能检测方法

Intelligent detection method of image water level inversion for water level without water scale in complex scenes

实现精细化水务管控和洪涝灾害预警,需要实时、准确感知水位突变事件。现有技术不能满足夜晚、雾霾、雨天、雪天、漂浮物遮挡及阴影等复杂恶劣环境下的水位识别需求。为此,本文提出一种融合改进YOLOv5与卡尔曼滤波原理的无水尺水位智能检测技术:①引入YOLOv5对水位线(水岸分界线)进行检测,并利用线性拟合方法获得实际水位线;②针对水位线在延伸方向无限大而在其法向无限小特点,提出强化中尺度特征的多层级特征融合方法改进原YOLOv5算法;③利用卡尔曼滤波引入水位历史信息作为先验知识,提高本技术对复杂恶劣环境的泛化性能;④将图像中事先标定的固定的标志物加入到深度学习网络中训练,根据标志位真实尺寸解算实际水位高程,实现无水尺检测方案。相关试验和实践表明,改进的YOLOv5更加轻量化;本文所述水位智能检测技术斜率准确性为97.3%,较原算法提高了2.4%;截距准确性为99.3%,较原算法提高了0.5%;在夜晚、雾霾、雨天、雪天、漂浮物遮挡及阴影等复杂恶劣环境下可以自动、准确识别出水位高程,误差小于0.1 m。

Realizing fine water control and flood warning requires real-time and accurate perception of sudden water level change events.The existing water level recognition technology cannot meet the needs of water level recognition in complex and harsh environments such as night,haze,rain,snow,floating object occlusion and shadow.To this end,this paper proposed an intelligent water level detection technique without water scale by integrating improved YOLOv5 and Kalman filtering principles:①Introducing YOLOv5 to detect the water level line(water shore demarcation line)and using linear fitting methods to obtain the actual water level line;②The water level is infinitely large in the extension direction and infinitely small in its normal direction.,a multi-level feature fusion method was proposed to strengthen the mesoscale features to improve the original YOLOv5 algorithm;③Using Kalman filtering to introduce water level history information as a priori knowledge to improve the generalization performance of this technique to complex and harsh environments;④Adding a fixed marker pre-calibrated in the image to the deep learning network for training,and solving the actual water level elevation based on the real size of the marker to achieve a water-rule-free detection scheme.Relevant experiments and practice showed that the improved YOLOv5 was more lightweight;the slope accuracy of the water level intelligent detection method described in this paper was 97.3%,which was 2.4%higher than the original model;the intercept accuracy was 99.3%,which was 0.5%higher than the original model;the water level elevation could be automatically and accurately identified in complex and harsh environments such as night,haze,rain,snow,floating object occlusion,shadow,and the error was less than 0.1 m.