Abstract: 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: ① Introducing YOLOv5 to detect the water level line (water shore demarcation line) and using linear fitting methods to obtain the actual water level line; ② The 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; ③ Using 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; ④ Adding 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 0.1m.

Key words: water level identification, water level detection without water gauge, deep learning, YOLOv5, Kalman filtering