CN-121982652-A - Method and device for detecting overflow of pipe well

CN121982652ACN 121982652 ACN121982652 ACN 121982652ACN-121982652-A

Abstract

The invention relates to a method and a device for detecting overflow of a piping shaft, and belongs to the field of urban public safety monitoring. The method comprises the steps of collecting images covering the well lid and water body characteristics, preprocessing and marking, adopting a YOLO26s target detection model to carry out recognition training, taking frames of a real-time video stream according to frequency f, sequentially inputting the obtained images into the trained YOLO26s model, outputting detection results of various characteristics in each frame, carrying out space-time combined verification through characteristic overlapping degree and multi-frame statistics, distinguishing vehicle splash and real overflow, setting a sliding frame number window, counting the number of frames of each characteristic in the window, and judging the triggering early warning level according to priority order. The invention not only greatly reduces false alarm rate of overflow detection, but also can realize multi-stage early warning from occurrence to explosion, and improves early warning speed.

Inventors

GONG WEIHAO
FANG AIYIN
YIN XIMENG
YAN BINGYANG
YUAN KANG
CUI MIAOMIAO

Assignees

山东锋士信息技术有限公司

Dates

Publication Date: 20260505
Application Date: 20260316

Claims (10)

1. The overflow detection method for the pipe well is characterized by comprising the following steps of: S1, acquiring an image covering the well lid and the water body characteristics; s2, preprocessing the image and labeling the visual detection class characteristics and the state judgment class characteristics; s3, performing recognition training by using a YOLO26s target detection model, performing dynamic characteristic enhancement on the injection time sequence differential noise of the input image in a training stage, and optimizing by using a loss function; S4, taking frames of the real-time video stream according to the frequency f, sequentially inputting the obtained images into a trained YOLO26s model, and outputting detection results of various features in each frame, wherein the detection results comprise boundary frame coordinates, state categories and confidence degrees; s5, performing space-time joint verification through feature overlapping degree and multi-frame statistics to distinguish vehicle splash and real overflow; and S6, setting a sliding frame number window, counting the number of frames of each feature in the window, and judging the triggering early warning level according to the priority order.
2. The method of claim 1, wherein the water features in step S1 include waves, water bloom, water column, and water accumulation.
3. The method for detecting overflow of a piping well according to claim 1, wherein step S2 includes 5 kinds of visual detection type features and 3 kinds of status determination type features, and the labeling specification is: (1) Visual inspection class characteristics: ① Marking a boundary frame which completely covers the edge of the well cover; ② Marking a texture area with the surface of the well lid being concentric circle diffusion, wherein the texture area is generated by water flow disturbance; ③ Marking water flow tracks which are sprayed out from the hole seams of the well cover and are scattered, and forming the water flow tracks irregularly; ④ Marking a region around the well lid with specular reflection characteristics or water surface textures, wherein the boundary frame covers more than 80% of the region for water accumulation; ⑤ The water column marks the water flow form which is vertically or obliquely upwards gushed from the hole seam of the well cover and is in a continuous column shape or a bundle shape, and is characterized by having a certain height, being relatively regular in form and being different from scattered water flowers; (2) Status determination class characteristics: ① The well lid is visible, namely the well lid is defined as being judged when the detection confidence of the model to the class of the well lid is larger than a preset threshold value; ② The well lid disappears, namely, judging when the detection confidence of the model to the class of the well lid is lower than a preset threshold value; ③ And (3) well lid displacement, namely calculating Euclidean distance offset of the center coordinates of the current well lid and the center coordinates of the historical reference frame on the premise that the well lid is visible, and judging if the offset is greater than a certain proportion of the width of the reference boundary frame.
4. The method for detecting overflow of a piping well according to claim 1, wherein the time sequence differential noise in the step S3 is to enhance the generalization ability of the model in a rain and fog environment by simulating the dynamic form of the surface corrugation of the well lid in the rain and water falling process.
5. The method for detecting overflow of a piping well according to claim 1, wherein the Loss function in the step S3 optimizes regression accuracy of a prediction frame by adopting CIoU Loss Loss functions, ensures accurate positioning of a well lid and a water accumulation area, and simultaneously balances foreground and background samples by combining a Focal Loss function, thereby relieving training deviation caused by rarity of ripple and water bloom small target samples.
6. The method for detecting overflow of a piping well according to claim 1, wherein the step S5 of distinguishing between splash and true overflow of a vehicle comprises: (1) Space positioning constraint, namely only counting water body characteristics in a well lid boundary frame, and requiring IOU (input output unit) of a water accumulation frame and a well lid reference frame to be more than 0.5, and eliminating road surface interference away from a well lid; (2) And (3) verifying the time duration, namely, only taking a single frame of water spray as a candidate, upgrading the water spray at the same position into a water column only when a plurality of continuous frames detect the water spray, and filtering out instantaneous water spray caused by vehicle water spray by adopting a sliding window statistical feature accumulated frame number.
7. The method for detecting overflow of a piping well according to claim 1, wherein the pre-warning in step S6 automatically determines 3 levels according to priority: (1) Early warning, namely, the manhole cover disappears but the water body features are insufficient, or slight ponding occurs when the manhole cover is visible; (2) The moderate risk is that the well lid is visible and obvious water body characteristics appear, or intermittent water columns appear; (3) Emergency event, well lid displacement, or well lid disappearance with obvious water body characteristics, or continuous water column appearance.
8. The method for detecting overflow of a piping well according to claim 1, wherein the pre-warning specific determination method in step S6 is as follows: (1) Preferentially, judging according to the characteristics of the water column: ① If the number of frames of the water column characteristics in the window is greater than a threshold T 1 , triggering an emergency level 3, and indicating that the spraying is continuous rather than accidental sputtering; ② If the number of frames of the water column characteristics in the window is smaller than T 1 and larger than an anti-false alarm threshold T 2 , triggering a moderate risk level 2, and indicating intermittent gushing or unstable pressure; (2) And then judging according to the well lid characteristics, when the number of frames of 'well lid disappearance' in the window is greater than the number of frames of 'well lid visibility': ① If the sum of the frame numbers of any feature of the ponding, the water bloom and the ripple is larger than a threshold T 3 , triggering an emergency level 3 to indicate serious ponding; ② If the sum of the frame numbers of any feature of water accumulation, water bloom and corrugation is smaller than T 3 , triggering an early warning level 1, and indicating that the well lid is possibly blocked by non-water sundries; (3) When the number of frames of the 'well lid disappearance' in the window is smaller than the number of frames of the 'well lid visible': ① If the proportion of the frame number of the 'well lid displacement' appearing in the window to the frame number of the 'well lid visible' exceeds a preset proportion threshold value P, triggering an emergency level 3 to indicate that the well lid is displaced; ② If the sum of the frame numbers of any feature of the ponding, the ripple and the water bloom of IOU >0.8 is larger than a threshold T 3 , triggering a moderate risk level 2 to indicate that the ponding is moderate; ③ If the sum of the frame numbers of any feature of 'ponding', 'ripple', 'water bloom' of 0.5< IOU <0.8 is larger than a threshold T 3 , triggering an early warning level 1, and indicating slight ponding; the early warning state is maintained for at least 10 seconds, if a higher grade appears in the early warning state, the early warning state is updated, otherwise, the current grade is maintained, and when all the conditions are not met, the early warning of any grade is not triggered.
9. The overflow detection device for the pipe well is characterized by comprising a data acquisition and preprocessing unit, a YOLO26s target detection model construction and detection module, a space-time joint verification module, a sliding window grading statistics module and an early warning grade judgment module, The data acquisition and preprocessing unit is used for deploying a camera with a fixed visual angle at a position 3-6 meters above a well cover which is easy to overflow, acquiring images of well covers and water body characteristics covering all environmental scenes, and preprocessing the images; The method comprises the steps that a YOLO26s target detection model construction and detection module is used for carrying out recognition training by adopting the YOLO26s target detection model, frames are taken from a real-time video stream according to frequency f, the obtained images are sequentially input into the trained YOLO26s model, and detection results of various features in each frame are output, wherein the detection results comprise boundary frame coordinates, state categories and confidence level; the space-time joint verification module is used for carrying out space-time joint verification through feature overlapping degree and multi-frame statistics to distinguish vehicle splash and real overflow; The sliding window grading statistics module is used for setting a sliding frame number window and counting the number of frames of each feature in the window; And the early warning level judging module judges the triggering early warning level according to the priority order.
10. The apparatus of claim 1, wherein the alert level determination module automatically determines 3 levels by priority based on feature accumulation within a sliding time window: (1) Early warning, namely, the manhole cover disappears but the water body features are insufficient, or slight ponding occurs when the manhole cover is visible; (2) The moderate risk is that the well lid is visible and obvious water body characteristics appear, or intermittent water columns appear; (3) Emergency event, well lid displacement, or well lid disappearance with obvious water body characteristics, or continuous water column appearance.

Description

Method and device for detecting overflow of pipe well Technical Field The invention relates to a method and a device for detecting overflow of a piping shaft, and belongs to the field of urban public safety monitoring. Background The traditional method for detecting the well lid problem mainly relies on manual inspection and fixed monitoring equipment, is relatively visual, but requires a large amount of manpower and time, cannot monitor all the weather generally, is easily interfered by human factors, and has low efficiency. In recent years, with the vigorous development of computer vision technology, well lid defect detection by deep learning is a new detection method. Among them, YOLO series networks are widely used because of their unique design and excellent performance, which are superior in image defect detection. The patent CN 120655610A discloses a well lid defect detection method and system based on an improved YOLOv n model, wherein the method comprises the steps of obtaining a well lid defect image dataset, dividing the well lid defect image dataset into a training set, a verification set and a test set, constructing an improved YOLOv n model introducing a multi-scale convolution, a channel attention mechanism and a residual structure, training the improved YOLOv n model based on the training set, the verification set and the test set to obtain a trained improved YOLOv8n model, and inputting a well lid defect image to be detected into the improved YOLOv n model for recognition to obtain a detection result. The patent CN 119672418A discloses a method and a system for integrated identification of hidden danger of a well lid by fusing target detection and segmentation. And extracting multiscale characteristics of the pretreated well lid picture by YOLOv x, and outputting a well lid prediction result. And adopting a Bagging integrated learning strategy, and obtaining a discrimination result of the hidden danger category of the well lid by using a weighted voting mechanism based on the well lid prediction result. And then inputting the preprocessed well lid picture into an encoder module consisting of PVTv-B4 for well ring feature extraction, and outputting a well ring prediction result. And simultaneously, carrying out secondary analysis on the well ring based on the well ring prediction result to obtain a discrimination result of the well ring hidden danger class. Finally, the hidden danger category is obtained by combining the distinguishing result of the well lid and the well ring. The technology is only based on the physical defect of a single-frame image detection well lid, can not utilize multi-frame information to eliminate instantaneous interference (such as vehicle splash and raindrops), can not realize the prejudgment of overflow trend, is easily influenced by shielding and illumination change, and has weak robustness. Disclosure of Invention The invention aims to overcome the defects and provide a method and a device for detecting overflow of a pipe well, so as to solve the problems that simple image identification in the prior art is difficult to distinguish vehicle splash from real overflow and has high false alarm rate, and a multistage early warning mechanism of the whole process from occurrence to explosion of overflow of a well cover is lacked. The technical scheme adopted by the invention is as follows: a method for detecting overflow of a pipe well comprises the following steps: S1, acquiring an image covering the well lid and the water body characteristics; s2, preprocessing the image and labeling the visual detection class characteristics and the state judgment class characteristics; s3, performing recognition training by using a YOLO26s target detection model, performing dynamic characteristic enhancement on the injection time sequence differential noise of the input image in a training stage, and optimizing by using a loss function; S4, taking frames of the real-time video stream according to the frequency f, sequentially inputting the obtained images into a trained YOLO26s model, and outputting detection results of various features in each frame, wherein the detection results comprise boundary frame coordinates, state categories and confidence degrees; s5, performing space-time joint verification through feature overlapping degree and multi-frame statistics to distinguish vehicle splash and real overflow; and S6, setting a sliding frame number window, counting the number of frames of each feature in the window, and judging the triggering early warning level according to the priority order. In the above method, the water body features in step S1 include waves, water bloom, water column and water accumulation. Step S2 comprises 5 types of visual detection type features (model needs to be learned) and 3 types of state judgment type features (post-processing logic generation), and the labeling specification is as follows: (1) Visual inspection class characteristics: ① Marking a b