Method and system for detecting methanol vehicle in sewage treatment plant area

Abstract

The invention discloses a method and a system for detecting methanol vehicles in a sewage treatment plant area. The method comprises the following steps: S1, acquiring video data in different areas of the sewage treatment plant area by adopting an acquisition strategy combining spatial subarea and time dynamic frame rate; s2, features weakly related to the vehicle state are removed from each frame of data through a mutual information algorithm, and a data set is expanded by adopting a data enhancement method; s3, constructing an improved model based on YOLOv8, and performing training by using the data set; s4, inputting real-time data collected in a certain area of the sewage treatment plant area into the trained improved model, and outputting vehicle positions, categories and confidence coefficients; and S5, a strategy of combining multi-frame verification and composite conditions is adopted to judge the vehicle state, the vehicle state comprises vehicle existence, vehicle leaving and same vehicle re-identification, and an anti-shake strategy is adopted to carry out voice early warning on a judgment result. According to the invention, the special scene requirement of methanol vehicle monitoring in a sewage treatment plant area is met.

Description

TECHNICAL FIELD

[0001] The present application relates to sewage treatment plant area methanol vehicle monitoring technology, and particularly relates to a sewage treatment plant area methanol vehicle detection method and system based on an improved YOLOv8 algorithm. BACKGROUND

[0002] In the sewage treatment plant area, real-time detection and state tracking of specific vehicles (such as methanol vehicles and other dangerous chemical transport vehicles) are key technologies to ensure the safety of the plant area. In the early days, such monitoring relied on manual visual identification, which was not only time-consuming and labor-intensive, but also prone to missed detection and false detection due to factors such as fatigue and lack of concentration.

[0003] With the development of computer vision and deep learning, vehicle detection methods based on target detection algorithms (such as YOLO and Faster R-CNN) have gradually become popular, significantly improving detection efficiency. However, the present inventors have found that these detection methods still face many challenges in actual scenarios:

[0004] (1) Low accuracy of state judgment: Traditional methods ignore the continuity of vehicle movement and multi-frame correlation information, resulting in large deviations in judging states such as "vehicle leaving" and "same vehicle reappearing". This can lead to problems such as "temporary occlusion" misjudgment as "leaving" or "slow departure not identified in time", resulting in high false positive and false negative rates and affecting monitoring reliability.

[0005] (2) Poor interactive experience: Voice prompts lack a jitter prevention mechanism, and frequent triggering of prompts for the same state (such as the continuous presence of a vehicle) can cause auditory interference and even lead to a decrease in the monitoring personnel's sensitivity to key alarm information.

[0006] (3) Insufficient performance and scene adaptability: Processing each frame of image in pursuit of accuracy results in low frame rate (<10FPS), which cannot meet the real-time monitoring requirements; or sacrificing detection stability to improve frame rate significantly increases the cost of manual review in complex scenarios (such as rainy days and night).

[0007] In addition, the present inventors have found that traditional state tracking techniques often use "fixed time threshold" or "continuous frame counting" to judge vehicle state (presence / leaving), lacking quantitative analysis of vehicle movement trends. The movement characteristics of vehicles in different scenarios (such as driving speed and dwell time) differ significantly, but traditional techniques use a unified decision standard, which cannot distinguish between "slow departure" and "temporary occlusion", resulting in state judgment bias. SUMMARY

[0008] The present application aims to provide a sewage treatment plant area methanol vehicle detection method and system based on an improved YOLOv8 algorithm to meet the special scenarios (such as high dust, scattered areas, and high safety requirements) of sewage treatment plant area methanol vehicle monitoring.

[0009] To this end, the present application provides a sewage treatment plant area methanol vehicle detection method, comprising the following steps: S1, using a space division region and time dynamic frame rate combined acquisition strategy to collect video data in different regions of the sewage treatment plant area; S2, eliminating features weakly related to the vehicle state from each frame of data through a mutual information algorithm, retaining multi-dimensional features strongly related to the vehicle state, and expanding the data set using a data enhancement method; S3, constructing an improved model based on YOLOv8 and training it using the data set, which includes embedding a coordinate attention module in the Backbone layer of the original model, adding a small target enhancement branch module to the Neck layer, and using a dynamic confidence threshold module in the Head layer; S4, inputting real-time data collected in a certain region of the sewage treatment plant area into the trained improved model, and outputting the vehicle position, category and confidence; S5, using a multi-frame verification and compound condition combined strategy to determine the vehicle state, including vehicle presence, vehicle departure and same vehicle re-identification, and using a de-bouncing strategy to give a voice warning.

[0010] The present application also provides a sewage treatment plant area methanol vehicle detection system which stores a computer program, which when executed, is used to implement the steps of the above sewage treatment plant area methanol vehicle detection method.

[0011] Compared with the prior art, the present application has the following technical effects:

[0012] 1) Improve vehicle state determination accuracy: combine detection box movement distance quantification (such as the distance between the center points of the initial frame and the current frame / the diagonal length of the initial frame) with a continuous disappearance frame threshold (such as 10 frames) to construct a compound determination mechanism, solve the one-sidedness of determining vehicle "existence / departure" only by a fixed time threshold, and accurately identify complex scenarios such as slow departure and temporary occlusion of vehicles.

[0013] 2) Realize same vehicle re-identification: set a re-appearance distance threshold (such as 0.15), compare the movement ratio of the last valid detection frame before the vehicle disappears and the detection frame when it reappears, solve the problem of misjudgment as a "new vehicle" after temporary occlusion of the vehicle, and avoid repeated alarms.

[0014] 3) Optimize voice prompt logic: introduce a voice cooling mechanism (such as a 3-second interval), control the voice prompt for the same state (such as continuous vehicle presence) to prevent shaking, and solve the problem of auditory interference caused by frequent repeated prompts.

[0015] 4) Solve the problem of single-frame detection fluctuation: By constructing a multi-frame detection queue (such as a 15-frame sliding window), the detection results of consecutive frames are fused and the effective ratio is calculated to avoid fluctuations in detection results caused by blurring of single-frame images and changes in illumination, thus providing stable basic data for vehicle status judgment.

[0016] 5) Balancing detection performance and real-time performance: By using a frame skipping strategy (such as processing 1 frame every 8 frames), the computational load is reduced while ensuring detection accuracy, resolving the contradiction between "low frame rate for full frame processing" and "sacrificing accuracy to increase frame rate", and ensuring stable operation at the frame rate required for real-time monitoring (such as above 25 FPS).

[0017] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0019] Figure 1 This is a flowchart of the method for detecting methanol vehicles in a wastewater treatment plant area according to the present invention;

[0020] Figure 2 This is an architecture diagram of the improved YOLOv8 model of this invention;

[0021] Figure 3 This is a block diagram of the methanol vehicle detection system for wastewater treatment plants according to the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] Combined with reference Figure 1 and Figure 2 The method for detecting methanol vehicles in wastewater treatment plant areas according to the present invention includes the following steps:

[0024] S1, Multi-scene video data acquisition

[0025] According to the monitoring needs of different areas of the sewage treatment plant, a "spatially divided area + time dynamic frame rate" collection strategy is adopted. Spatially, three key areas: warehouse entrances and exits for fine monitoring of loading and unloading processes, passageways for tracking driving trajectories, and plant entrances for recording overall entry and exit times. Temporally, the dynamic frame rate adjustment algorithm adapts to vehicle activity intensity: automatically reduce the frame rate to 5FPS during no-vehicle periods to save resources, and increase it to 30FPS within 1 second when a vehicle is detected to capture details, ensuring real-time monitoring while reducing 70% of invalid data collection.

[0026] S2, Vehicle Feature Engineering

[0027] Through multi-dimensional feature screening and enhancement, the adaptability of the model to complex environments is improved. Core features include the center point coordinates of the vehicle detection box, the aspect ratio, the displacement of the last three frames, and the speed. Weakly related features (such as background brightness, MI<0.2) are removed through mutual information algorithms, and strongly related features (such as displacement, MI>0.6) are retained. The data enhancement section simulates the sewage treatment plant scene: adds Gaussian noise (σ=0.02) to simulate dust pollution, generates 5-15 pixel rain lines to simulate rainy weather, and performs ±2 second time translation on historical trajectory data to expand the training set, improving the detection stability of the model in bad weather by 50%.

[0028] S3, Building an Improved Model Based on YOLOv8

[0029] YOLOv8 base model features: 1. Backbone layer uses CSPDarknet structure, extracts multi-dimensional features through cross-stage connection, balances feature retention and computational efficiency; 2. Neck layer uses PAN-FPN to realize multi-scale feature fusion, balances different size target detection capabilities; 3. Head layer uses anchor-free design, directly predicts target position and class, simplifies model structure; 4. After inference, use fixed confidence threshold (default 0.25) and NMS (IoU=0.45) to filter results, fast processing speed, suitable for real-time scenarios, but insufficient in capturing small targets and outline features in complex environments.

[0030] The improved model based on YOLOv8 is the core of improving detection stability, and optimizes three points for the sewage treatment plant scene, such as Figure 2As shown: the Backbone layer adds a coordinate attention module, which separately weights the x and y direction features to enhance the extraction ability of vehicle contour features; the Neck layer adds a small target branch to separately convolve the distant vehicles with <32x32 pixels, solving the problem of missing detection of distant vehicles in the warehouse area; the Head layer uses a dynamic confidence threshold (0.5~0.7) to automatically adjust according to the image clarity, reducing the false detection rate by 40% in complex environments. After reasoning, the non-maximum suppression (NMS, IoU=0.5) is used to remove duplicate frames, and the accurate vehicle position and category are output.

[0031] S4, model inference

[0032] The real-time data collected in a certain area of the sewage treatment plant is input into the improved model, and the vehicle position, category and confidence are output.

[0033] Before inputting the collected real-time data into the improved model, frame skipping processing is performed to reduce the number of processing frames. Then the adaptive bilateral filtering algorithm is used to dynamically adjust the filtering parameters, and the Retinex algorithm is used to enhance the vehicle contour in backlit environment.

[0034] After the improved model outputs the results, the detection frame pixel coordinates are uniformly converted into relative image size proportion values, and the first frame number sliding window is used to weight and smooth the detection results output by the model. The average value of each frame detection frame coordinates is calculated according to the principle of high recent frame weight and low long-term frame weight, so that multiple frames are fused as one frame to determine the vehicle state.

[0035] S5, vehicle state judgment and monitoring and early warning mechanism

[0036] The vehicle state is determined by "multi-frame verification + compound condition", and the core is the whole process logic of "existence - leaving - re-identification". Existence state determination: when the proportion of "vehicle" frames in the 15-frame sliding window is ≥70%, and the vehicle is detected for 3 consecutive frames, it is determined that "vehicle exists", and the initial detection frame (B0) is recorded as the position reference. Each frame here specifically refers to each frame after fusion. The leaving state determination adopts double conditions: one is to calculate the moving proportion (R) of the current frame (B i ) and B0, the formula is:

[0037]

[0038] Where R is the moving proportion (output, range 0~1), the larger the value, the farther the movement, is the center point coordinate of the current vehicle detection frame, is the center point coordinate of the initial vehicle detection frame, These are the coordinates of the top left and bottom right corners of the initial vehicle detection frame (used to calculate the diagonal length of the initial frame).

[0039] When R > 30%, it is marked as "about to leave"; secondly, the number of consecutive frames in which no vehicle is detected is counted, and when it is ≥ 10 frames and the duration is > 2 seconds, it is judged as "completely left".

[0040] The same vehicle re-identification is achieved through position similarity matching: the last valid bounding box (B_last) before the vehicle disappears is recorded. When the vehicle is detected again, the movement ratio between the new bounding box and B_last is calculated. If it is less than 15%, it is determined to be the same vehicle (no repeated alarm); otherwise, it is determined to be a new vehicle (alarm is triggered), which solves the problem of repeated alarms after occlusion (accuracy rate reaches 95%).

[0041] A tiered early warning system has been established, with voice prompts strictly adhering to anti-shake rules: Level 1 warning (first entry of a vehicle) broadcasts "Methanol vehicle has arrived"; Level 2 warning (staying for more than 30 minutes) broadcasts "Vehicle stay exceeds time limit, please be aware"; and Level 3 warning (departure) broadcasts "Methanol vehicle has left." By recording the time and content of the last prompt, the system ensures that the playback interval of the same warning content is ≥3 seconds, avoiding "repeated prompts when a vehicle is continuously present." This reduces invalid voice prompts in the monitoring room by 80%, and simultaneously links with the plant's security equipment to achieve tiered control, improving safety management efficiency.

[0042] In practical deployment, the detection method based on the present invention also includes a performance optimization step (solving the balance between performance and real-time performance).

[0043] The TensorRT program was used to accelerate the model, reducing the inference time from 120ms to 45ms (GPU: GTX1650), ensuring stable operation on the industrial control host (CPU utilization <30%). Ultimately, a balance was achieved between "96% detection accuracy + 25FPS real-time frame rate," meeting the 24 / 7 uninterrupted monitoring requirements of the wastewater treatment plant.

[0044] To address the specific challenges of monitoring methanol vehicles in wastewater treatment plant areas (such as high dust levels, dispersed locations, and stringent safety requirements), this invention also provides a methanol vehicle detection system for wastewater treatment plant areas, such as... Figure 3 As shown, it includes a data acquisition module, a data processing module, an improved model based on YOLOv8, and a monitoring and early warning module.

[0045] The data acquisition module uses video data from a monitoring point in the wastewater treatment area and balances accuracy and speed through resource scheduling strategies. For example, it performs frame skipping on the video stream, performing inference once every 8 frames, reducing the computational load by 7 / 8.

[0046] The data processing module is a key component in addressing the volatility of single-frame detection. It eliminates environmental interference through a two-stage process: adaptive preprocessing before model inference and multi-frame fusion after model inference. Simultaneously, it optimizes computational resource allocation to balance performance and real-time capabilities. Specifically, it includes:

[0047] Adaptive preprocessing (before model inference):

[0048] To address dust pollution in wastewater treatment plant areas, an adaptive bilateral filtering algorithm is used to dynamically adjust the filtering parameters (σ value adaptively varies with image noise level between 0.1 and 0.5). Simultaneously, the Retinex algorithm is used to enhance vehicle outlines under backlight conditions. The formula is as follows:

[0049]

[0050] in, This is the enhanced image (output result). The original image in coordinates Pixel value at that location, It is a Gaussian kernel function (used to estimate the illumination components of an image). It's a convolution operation. This is a logarithmic transformation used to compress the dynamic range of an image. This operation improves the detection confidence in low-light scenes to above 0.6.

[0051] Multi-frame fusion post-processing (after model inference):

[0052] The format conversion process uniformly converts the pixel coordinates of the detection box into proportional values ​​relative to the image size, thus eliminating resolution differences between different cameras.

[0053] During real-time data cleaning, a 15-frame sliding window is used to weight and smooth the detection results output by the model. The average value of the detection box coordinates in each frame is calculated according to the principle of "higher weight for recent frames and lower weight for older frames", as shown in the following formula:

[0054]

[0055] in, This is the smoothed vehicle coordinates (output result). It is the first Vehicle detection bounding box coordinates of the frame (input, = 1, 2, ..., 15), It is the first The weight of frames is assigned, with more recent frames having a higher weight (e.g., the weight of frame 15 is (16-15) / 120 = 1 / 120, and the weight of frame 1 is 15 / 120), and the total weight is 1. This operation significantly reduces single-frame jitter caused by rain and dust (the fluctuation amplitude is reduced by 60%).

[0056] The data storage adopts a "hot caching + cold storage" strategy: the most recent 300 frames of data are stored in a circular queue for multi-frame analysis, and historical data are partitioned by "region-date" and stored in a time-series database. The storage interval is 1 second / record for periods with vehicles and extended to 10 seconds / record for periods without vehicles, reducing storage resource usage by 60%.

[0057] The improved YOLOv8-based model uses an improved YOLOv8 as its core, enhancing state judgment accuracy and re-identification capabilities through network structure optimization and multi-dimensional algorithms. For the specific characteristics of vehicle detection in wastewater treatment plant areas, the improvements to YOLOv8 include:

[0058] 1. The backbone layer embeds a coordinate attention module, which enhances the feature extraction capability for partially occluded vehicles by separately weighting the vehicle position features (x and y directions). The formula is as follows:

[0059]

[0060] in, These are coordinate attention weights (output, range 0~1). It is the feature map in Pixel value at that location, These are the height and width of the feature map. It is a convolution operation (used to compress feature dimensions). It is the sigmoid activation function (which normalizes the weights to 0~1).

[0061] 2. Improves small target detection recall by 40%; A new small target enhancement branch is added to the Neck layer, which performs separate convolution processing on distant vehicle features < 32×32 pixels to solve the problem of missed detection of distant vehicles in warehouse areas.

[0062] 3. The Head layer uses a dynamic confidence threshold, which is automatically adjusted based on image sharpness (Laplacian operator score). The formula is as follows: .in, It is the dynamic confidence threshold (output, range 0.5~0.7). This is the image sharpness score (input, calculated based on the Laplacian operator, ranging from 0 to 1), with higher values ​​indicating sharper images. The threshold is lowered to reduce missed detections when dust pollution is severe, and raised to reduce false detections on sunny days.

[0063] The state analysis submodule determines the vehicle's departure state through a composite judgment mechanism of "movement distance quantification + consecutive disappearance frame count". At the same time, it realizes the re-identification of the same vehicle based on position similarity matching, which completely solves the one-sidedness of traditional single threshold judgment.

[0064] The monitoring and early warning module focuses on the practicality of monitoring the wastewater treatment plant area, optimizing voice prompts through tiered early warning and anti-shake mechanisms. The monitoring interface supports overlay display of the plant area map, marking vehicle location, status, and dwell time in real time. An automatic flashing alarm is triggered when a vehicle remains in the warehouse area for more than 30 minutes. The voice method employs a dual anti-shake mechanism of "content + time": it records the content and time of the last prompt; if the current prompt content is the same and the interval is less than 3 seconds, it will not be repeated; voice prompts are only triggered when the content changes or the interval is ≥3 seconds, avoiding auditory interference in the monitoring room. Simultaneously, it links the plant area gate and audible / visual alarms via a PLC interface to achieve closed-loop control of "automatically closing the gate and activating a high-decibel alarm when a vehicle is abnormally detained."

[0065] The technical effects of this invention are reflected in the following aspects:

[0066] 1. Improved detection accuracy

[0067] By employing multi-frame fusion algorithms and motion trend analysis, the accuracy of vehicle status determination is significantly improved. For common scenarios in wastewater treatment plant areas, such as dust obstruction and backlighting, the method increases the accuracy of determining the presence / departure of methanol vehicles to 96%, and achieves a re-identification rate of 95% for the same vehicle. This effectively reduces false alarms (such as misjudging departure due to brief obstruction) and missed alarms (such as failing to identify slow departure), providing a reliable basis for the precise management of hazardous chemical vehicles in wastewater treatment plant areas and ensuring plant production safety, as detailed in Table 1 below.

[0068] Table 1. Comparison of Detection Accuracy

[0069]

[0070] 2. Enhanced real-time response

[0071] Employing a dynamic frame rate adjustment and rapid re-identification mechanism, the system can capture vehicle trajectory and status changes in real time. When a vehicle is suddenly obstructed (e.g., by a forklift for 5 seconds) or moves across areas (e.g., from a warehouse area into a passageway), the method can complete a status update within 1 second, avoiding regulatory blind spots caused by response delays and reducing the risk of safety accidents such as hazardous chemical leaks, as detailed in Table 2 below.

[0072] Table 2. Real-time Response Comparison Table

[0073]

[0074] 3. Optimize operation and maintenance costs

[0075] Through intelligent frame skipping and voice anti-shake mechanisms, the method reduces resource consumption by 60% (CPU utilization drops from 80% to 30%) while ensuring monitoring effectiveness. At the same time, it reduces invalid voice prompts by 80%, reducing auditory fatigue and workload for monitoring personnel. The ability to dynamically adapt to cameras in different factory areas also reduces equipment replacement costs, as shown in Table 3.

[0076] Table 3. Comparison of Operation and Maintenance Costs

[0077]

[0078] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting methanol vehicles in a wastewater treatment plant area, characterized in that, Includes the following steps: S1. A data acquisition strategy combining spatial regionalization and temporal dynamic frame rate is adopted to collect video data from different areas of the wastewater treatment plant. S2. Use the mutual information algorithm to remove features that are weakly correlated with the vehicle state from each frame of data, retain multi-dimensional features that are strongly correlated with the vehicle state, and use data augmentation methods to expand the dataset. S3. Construct an improved model based on YOLOv8 and train it using the dataset. The improved model based on YOLOv8 includes: embedding a coordinate attention module in the Backbone layer of the original model; adding a small target enhancement branch module in the Neck layer; and using a dynamic confidence threshold module in the Head layer. S4. Input the real-time data collected from a certain area of ​​the sewage treatment plant into the trained improved model, and output the vehicle location, category and confidence level; S5. A strategy combining multi-frame verification and composite conditions is used to determine the vehicle status, including vehicle presence, vehicle departure, and re-identification of the same vehicle. The determination results are then used to provide voice warnings using an anti-shake strategy.

2. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, The acquisition strategy includes: spatially acquiring data from three key areas: warehouse entrances and exits for detailed monitoring of the loading and unloading process, passageways for tracking driving trajectories, and factory entrances for recording overall entry and exit times; temporally, adapting to vehicle activity intensity through a dynamic frame rate adjustment algorithm: automatically reducing the frame rate to the first frame rate during periods without vehicles to save resources, and increasing it to the second frame rate within 1 second when a vehicle is detected to capture details.

3. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, Multi-dimensional features strongly correlated with vehicle status include the center point coordinates of the vehicle detection box, aspect ratio, displacement and velocity over three consecutive frames.

4. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, Data augmentation methods were employed to expand the dataset, including adding Gaussian noise to simulate dust pollution, generating rain lines of 5-15 pixels to simulate rainy weather, and shifting historical trajectory data by ±2 seconds to expand the training set.

5. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, The coordinate attention module is used to separately weight features in the x and y directions; The small target enhancement branch module is used to perform separate convolution processing on distant vehicles smaller than 32×32 pixels; the dynamic confidence threshold module is used to automatically adjust the threshold according to the image sharpness.

6. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, In step S4, frame skipping is performed before the collected real-time data is input into the improved model to reduce the number of processing frames.

7. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 6, characterized in that, In step S4, after improving the model output, the pixel coordinates of the detection boxes are uniformly converted into a ratio relative to the image size, and the detection results output by the model are weighted and smoothed using a sliding window of the first frame number. The average value of the detection box coordinates of each frame is calculated according to the principle that the recent frame has a higher weight and the distant frame has a lower weight, so as to fuse multiple frames into one frame to determine the vehicle state.

8. The method for detecting methanol vehicles in a wastewater treatment plant area according to claim 1, characterized in that, The anti-shake strategy records the content and time of the last prompt. If the content of the current prompt is the same and the interval is less than 3 seconds, it will not be played again. A voice warning will only be triggered when the content changes or the interval is greater than or equal to 3 seconds.

9. The method for detecting methanol vehicles in a wastewater treatment plant area according to any one of claims 1 to 8, characterized in that, In step S5, the strategy of combining multi-frame verification and composite conditions to determine the vehicle state includes: When the percentage of frames containing vehicles in the second frame sliding window is ≥70%, and vehicles are detected for 3 consecutive frames, it is determined that a vehicle exists, and the initial detection box B0 is recorded as the position reference. The departure state determination uses two conditions: first, calculate the current bounding box B. i The movement ratio R of B0 is used to mark the vehicle as about to leave when R > 30%; secondly, the number of consecutive frames in which no vehicle is detected is counted, and when it is greater than or equal to 10 frames and the duration is greater than 2 seconds, it is determined that the vehicle has completely left. The same vehicle re-identification is achieved through position similarity matching: record the last valid bounding box B_last before the vehicle disappears, and when the vehicle is detected again, calculate the movement ratio between the new bounding box and B_last. If it is less than 15%, it is determined to be the same vehicle; otherwise, it is determined to be a new vehicle.

10. A methanol vehicle detection system for a wastewater treatment plant, comprising a computer program, characterized in that, When executed, the program is used to implement the steps of the method for detecting methanol vehicles in a wastewater treatment plant area according to any one of claims 1 to 9.

Images