Sanitation event monitoring method and system based on multi-source data fusion and model cascade

Abstract

The application provides a sanitation event monitoring method and system based on multi-source data fusion and model cascade, comprising: collecting real-time operation data of a patrol vehicle, and performing AI identification on a sanitation risk vehicle to obtain corresponding parameters; performing Re-ID modeling on the patrol vehicle to obtain corresponding Re-ID feature information, and storing the Re-ID feature information and the parameters in the cloud; constructing a global sanitation risk vehicle dynamic topology graph according to the fixed time interval of uploading the cloud for the feature information and the parameters; collecting real-time traffic condition data, and performing data preliminary cleaning and processing on the real-time traffic condition data, and simultaneously constructing an AI multi-model cascade to generate a sanitation event prediction model according to the vehicle dynamic topology graph; and performing sanitation event monitoring through the model to obtain a sanitation risk area ranking. The application fuses multi-source heterogeneous full-quantity data to construct an AI multi-model cascade to generate a sanitation event prediction model, and improves the timeliness and effectiveness of sanitation event discovery.

Description

Technical Field

[0001] This invention relates to the field of sanitation operation incident processing technology, specifically to a sanitation incident monitoring method and system based on multi-source data fusion and model cascading. Background Technology

[0002] With the deepening of urbanization and the continuous development of urban scale, large-scale urban sanitation is an indispensable and important foundation for the normal operation of cities. However, due to the coupling and complexity of large-scale urban operations, a large number of sudden and sporadic sanitation events occur every day (such as large-scale garbage dumping on roads, accumulation of earthwork on roads, road dirt and disorder caused by traffic accidents, and industrial waste covering traffic auxiliary roads). These events need to be included in the sanitation event monitoring and prediction system and captured by the cloud platform as soon as possible.

[0003] Currently, the standard practice in the sanitation industry is to pre-deploy a large number of road inspection equipment and personnel to conduct routine inspections along predetermined routes in order to identify high-risk sanitation incidents as much as possible. However, since inspection routes are already determined by human experience, this requires a clear and accurate understanding of the overall road surface. If the accuracy is insufficient, the planned inspection routes may not cover all necessary suspicious areas, resulting in wasted manpower and resources in low-risk areas, and high-risk sanitation incidents may not receive timely warnings and monitoring. Moreover, given the constantly changing urban traffic flow, how to accurately identify and monitor high-risk vehicles (such as garbage trucks, dump trucks, and open-top trucks) that may cause sanitation incidents, analyze their past routes in real time, and then provide routes for high-risk areas that need to be inspected, and notify personnel and equipment currently inspecting along predetermined routes to adjust their routes, is a problem that cannot be effectively solved by manual experience-based scheduling.

[0004] Therefore, there is an urgent need in the market for a method and system for monitoring and predicting sanitation incidents that integrates multi-source heterogeneous full data. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for monitoring sanitation incidents based on multi-source data fusion and model cascading.

[0006] A method for monitoring sanitation incidents based on multi-source data fusion and model cascading, provided by the present invention, includes:

[0007] Step S1: Collect real-time operating data of inspection vehicles and use AI to identify sanitation risk vehicles to obtain corresponding parameters;

[0008] Step S2: Perform Re-ID modeling on the inspection vehicle to obtain the corresponding Re-ID feature information, and store the Re-ID feature information and the parameters in the cloud;

[0009] Step S3: Construct a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals.

[0010] Step S4: Collect real-time traffic data, perform preliminary cleaning and processing on the real-time traffic data, and simultaneously construct an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model.

[0011] Step S5: Monitor sanitation events using the sanitation event prediction model to obtain a ranking of sanitation risk areas.

[0012] Preferably, the parameters include real-time vehicle status parameters, high-risk vehicles, and static parameters of the high-risk vehicles;

[0013] The real-time vehicle status parameters include vehicle speed, driving trajectory, driving lane, driving direction, and driving posture;

[0014] High-risk vehicles refer to vehicles that could cause sanitation incidents, and are identified through a pre-defined whitelist.

[0015] The static parameters include vehicle type, license plate, and sanitation risk characteristics.

[0016] Preferably, the AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model;

[0017] The vehicle Re-ID model is used to associate vehicle information on the vehicle side and the drone side;

[0018] The state discrimination model is used to determine abnormal vehicle operation, which is related to sanitation incidents, including vehicle shaking and inadequate vehicle sealing.

[0019] The trajectory prediction model is used to identify the consistent trajectory of a vehicle by associating information from multiple sensors.

[0020] Preferably, step S2 includes:

[0021] Step S2.1: Read vehicle images from the sanitation cloud business image acquisition library and preprocess the vehicle images;

[0022] Step S2.2: Annotate the key points in the preprocessed vehicle image;

[0023] Step S2.3: By locating the key points and using the region segmentation method, the current vehicle is divided into four sides: front, back, left, and right.

[0024] Step S2.4: Generate corresponding region masks based on the four faces, and associate each region mask with the key point, wherein there is at least one pair of associations and they do not overlap.

[0025] Step S2.5: Extract region feature vectors from multiple region segmentation results using a CNN convolutional neural network;

[0026] Step S2.6: Fuse the region feature vector with the global feature vector;

[0027] Step S2.7: Perform vehicle re-identification and retrieval on the fused feature vector to obtain the corresponding feature information.

[0028] Preferably, step S5 includes matching the real-time vehicle status data monitored in the selected monitoring area with the collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the comprehensive data after matching. The calculation formula is as follows:

[0029]

[0030] Among them, a i This represents the current risk value of the vehicle, m. t P represents the weight percentage of road segments in the current time period. t This represents the risk prediction threshold set for the current time period, and f(t+Δt) represents the subsequent predicted risk coefficient for the current area.

[0031] According to the present invention, a sanitation incident monitoring system based on multi-source data fusion and model cascading includes:

[0032] Module M1: Collects real-time operating data of inspection vehicles and uses AI to identify sanitation risk vehicles to obtain corresponding parameters;

[0033] Module M2: Performs Re-ID modeling on the inspection vehicle to obtain the corresponding Re-ID feature information, and stores the Re-ID feature information and the parameters in the cloud;

[0034] Module M3: Constructs a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals.

[0035] Module M4: Collects real-time traffic data, performs preliminary cleaning and processing on the real-time traffic data, and constructs an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model;

[0036] Module M5: Monitors sanitation events using the sanitation event prediction model to obtain a ranking of sanitation risk areas.

[0037] Preferably, the parameters include real-time vehicle status parameters, high-risk vehicles, and static parameters of the high-risk vehicles;

[0038] The real-time vehicle status parameters include vehicle speed, driving trajectory, driving lane, driving direction, and driving posture;

[0039] High-risk vehicles refer to vehicles that could cause sanitation incidents, and are identified through a pre-defined whitelist.

[0040] The static parameters include vehicle type, license plate, and sanitation risk characteristics.

[0041] Preferably, the AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model;

[0042] The vehicle Re-ID model is used to associate vehicle information on the vehicle side and the drone side;

[0043] The state discrimination model is used to determine abnormal vehicle operation, which is related to sanitation incidents, including vehicle shaking and inadequate vehicle sealing.

[0044] The trajectory prediction model is used to identify the consistent trajectory of a vehicle by associating information from multiple sensors.

[0045] Preferably, module M2 includes:

[0046] Module M2.1: Reads vehicle images from the sanitation cloud business image acquisition library and preprocesses the vehicle images;

[0047] Module M2.2: Annotate key points in the preprocessed vehicle image;

[0048] Module M2.3: By locating the key points and using the region segmentation method, the current vehicle is divided into four sides: front, back, left, and right.

[0049] Module M2.4: Generate corresponding region masks based on the four faces, and associate each region mask with the key point, wherein there is at least one pair of such associations and they do not overlap.

[0050] Module M2.5: Uses a CNN convolutional neural network to extract region feature vectors from multiple region segmentation results;

[0051] Module M2.6: Merges the regional feature vector with the global feature vector;

[0052] Module M2.7: Performs vehicle re-identification and retrieval on the fused feature vector to obtain the corresponding feature information.

[0053] Preferably, module M5 includes: step S5 includes matching the real-time vehicle status data monitored in the selected monitoring area with the collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the comprehensive data after matching, using the following calculation formula:

[0054]

[0055] Among them, a i This represents the current risk value of the vehicle, m. t P represents the weight percentage of road segments in the current time period. t This represents the risk prediction threshold set for the current time period, and f(t+Δt) represents the subsequent predicted risk coefficient for the current area.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] This invention integrates AI modeling with multi-source heterogeneous full data, fully considers the impact of real-time traffic flow, identifies high-risk sanitation areas through AI algorithms, constructs a multi-model cascade AI model to generate a sanitation event prediction model, improves the timeliness and effectiveness of sanitation event detection, and enables intelligent near-real-time scheduling of human and material resources, thereby improving the accuracy and predictability of decision-making. Attached Figure Description

[0058] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0059] Figure 1 This is a schematic diagram of the workflow of the present invention.

[0060] Figure 2 This is a schematic diagram of the process of obtaining corresponding feature information through the Re-ID model in this invention.

[0061] Figure 3 This is a schematic diagram illustrating the effect of dividing a vehicle into four faces by locating key points and segmenting regions in this invention. Detailed Implementation

[0062] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0063] Example 1

[0064] According to the present invention, a method for monitoring sanitation incidents based on multi-source data fusion and model cascading is provided, such as... Figure 1 The following are included:

[0065] Step S1: Collect real-time operational data of inspection vehicles and use AI to identify high-risk sanitation vehicles, obtaining corresponding parameters. These parameters include real-time vehicle status parameters, high-risk vehicle parameters, and static parameters of high-risk vehicles. Real-time vehicle status parameters include vehicle speed, trajectory, lane, direction, and posture. High-risk vehicles are those that may cause sanitation incidents, such as garbage trucks, dump trucks, and open-top trucks, identified through a pre-defined whitelist. Static parameters include vehicle type, license plate, and sanitation risk characteristics.

[0066] Specifically, during the operation of patrol vehicles in road areas, the patrol vehicle group acts as the data collection source, collecting a large amount of real-time dynamic video frames / point cloud frame data. This data is then used to identify the real-time variable state parameters of the vehicles, i.e., the real-time vehicle status parameters, through AI algorithm models. Simultaneously, high-risk vehicles and their static parameters are also identified through the AI ​​model.

[0067] Step S2: Perform Re-ID modeling on the inspection vehicle to obtain the corresponding Re-ID feature information, and persistently store the Re-ID feature information and the parameters in the cloud. This step aims to capture accurate features of vehicles in different poses and backgrounds on the road that may cause sanitation incidents.

[0068] Specifically, such as Figure 2 As shown, step S2 includes:

[0069] Step S2.1: Read vehicle images from the sanitation cloud business image acquisition library and preprocess the vehicle images. The preprocessing includes exporting the vehicle images from the sanitation cloud business image acquisition library, filtering them by time, category, and image quality-related indicators during export.

[0070] Step S2.2: Annotate key points in the preprocessed vehicle images. Images of vehicles taken from different angles show different main content, thus expressing different local features. Therefore, it is necessary to annotate key points in the current images when creating the image training set to eliminate a large amount of noise during model training.

[0071] Step S2.3: By locating the key points and using the region segmentation method, the current vehicle is divided into four sides: front, back, left, and right. For example... Figure 3As shown, the vehicle is divided into four sides based on direction: front, back, left, and right. Each side focuses on the local description of multiple specific regions. This local description does not directly predict boundary points or corners, but rather selects key points as discriminant locations or key vehicle components, such as wheels, headlights, signs, rearview mirrors, and license plates. Ultimately, each vehicle image is marked with multiple key points, achieving segmentation of the target vehicle across multiple regions. Just as... Figure 3 Key points from number 1 to number 20.

[0072] Step S2.4: Generate corresponding region masks based on the four faces, and associate each region mask with the key point, wherein there is at least one pair of associations and they do not overlap.

[0073] Step S2.5: Use a CNN convolutional neural network to extract region feature vectors from the segmentation results of multiple regions.

[0074] Step S2.6: Fuse the regional feature vector with the global feature vector. First, perform fully connected processing on the local feature vector, and then perform softmax activation to obtain a set of weight coefficients. The weight coefficients and local features are weighted accordingly, and then connected to the global feature vector. Finally, the feature vector is flattened through a fully connected layer to obtain the fixed-format appearance feature vector output of the target vehicle.

[0075] Step S2.7: Perform vehicle re-identification and retrieval on the fused feature vector to obtain the corresponding feature information. The obtained vehicle appearance features can be directly compared with the vehicle appearance features in different vehicle images, solving the problem of indistinguishability between different regions of different vehicles.

[0076] Step S3: Construct a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals. This step aims to facilitate subsequent cascaded computation of multiple AI models.

[0077] Step S4: Collect real-time traffic data, perform preliminary data cleaning and processing on the real-time traffic data, and simultaneously construct an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model. The AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model. The vehicle Re-ID model is used to associate vehicle information from the vehicle side and the drone side. The state discrimination model is used to determine vehicle driving anomalies, which are related to the occurrence of sanitation events, including vehicle shaking and inadequate vehicle closure. The trajectory prediction model is used to identify the consistent trajectory of vehicles by associating information from multiple sensors.

[0078] Specifically, methods for collecting real-time traffic data include capturing real-time traffic data, such as video frames, point clouds, and vehicle conditions, through multiple drone swarms.

[0079] Step S5: Monitor sanitation incidents using the sanitation incident prediction model to obtain a ranking of sanitation risk areas. Step S5 includes matching the real-time vehicle status data monitored in the selected monitoring area with the collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the comprehensive data after matching. The calculation formula is as follows:

[0080]

[0081] Among them, a i This represents the current risk value of the vehicle, m. t P represents the weight percentage of road segments in the current time period. t This represents the risk prediction threshold set for the current time period, and f(t+Δt) represents the subsequent predicted risk coefficient for the current area.

[0082] It should be noted that the risk area ranking is based on the following data results for settlement, which include business processing data based on grid maps, vehicle identification and trajectory prediction data realized by vehicle Re-ID model, and data collected on static parameters of high-risk vehicles.

[0083] Furthermore, the following explains the processing of business data based on grid maps: Existing map data has relatively weak regional granularity, containing information on cities, districts, and streets. These areas are grid combinations with clear geographical connections, thus failing to meet the flexibility required for customized business operations. Therefore, a function for manually defining regions is provided. Users can freely divide the area range on the grid map by selecting boxes, and initialize a series of parameters for each area range, such as risk prediction thresholds, area monitoring time intervals, and the proportion of intersection and road inspections.

[0084] The vehicle Re-ID model for vehicle identification and trajectory prediction is explained as follows: Using the vehicle re-identification model trained in step S2, all vehicle images acquired from the vehicle acquisition source are used for discrimination. For identical vehicles, a spatiotemporal integrated storage method is employed to save the image state, geographical location, precise time, and real-time vehicle status data at the time of capture to the cloud. The cloud can then use the vehicle's current speed data, road congestion conditions, and road topology information to predict the vehicle's subsequent trajectory.

[0085] The following explains the collection of static parameters for high-risk vehicles: The drone swarm primarily captures abnormal sanitation events, such as multiple dump trucks suddenly appearing in a certain area, driving from a construction site to a waste treatment plant. This type of high-risk sanitation data is uploaded to the cloud by the drone swarm. The cloud then matches the returned data with the aforementioned vehicle Re-ID model to achieve spatiotemporal data for vehicle identification and trajectory prediction. After matching, the original spatiotemporal integrated data is enriched accordingly.

[0086] This invention aims to improve the timeliness and effectiveness of sanitation incident detection by fusing multiple data sources, combining real-time dynamic parameters of vehicles with static parameters of high-risk vehicles to dynamically update vehicle Re-ID modeling, and by cascading multiple AI models for sanitation risk vehicles—namely, vehicle Re-ID model, state judgment model, and trajectory prediction model—to derive a ranking of sanitation risk areas. This will enable intelligent, near-real-time scheduling of human and material resources, and improve the accuracy and predictability of decision-making.

[0087] Example 2

[0088] The present invention also provides a sanitation incident monitoring system based on multi-source data fusion and model cascading. Those skilled in the art can implement the sanitation incident monitoring system based on multi-source data fusion and model cascading by executing the steps of the sanitation incident monitoring method based on multi-source data fusion and model cascading. That is, the sanitation incident monitoring method based on multi-source data fusion and model cascading can be understood as a preferred embodiment of the sanitation incident monitoring system based on multi-source data fusion and model cascading.

[0089] According to the present invention, a sanitation incident monitoring system based on multi-source data fusion and model cascading includes:

[0090] Module M1: Collects real-time operational data of inspection vehicles and uses AI to identify high-risk sanitation vehicles, obtaining corresponding parameters. These parameters include real-time vehicle status parameters, high-risk vehicles, and their static parameters. Real-time vehicle status parameters include vehicle speed, trajectory, lane, direction, and posture. High-risk vehicles are those that could cause sanitation incidents and are identified through a pre-defined whitelist. Static parameters include vehicle type, license plate, and sanitation risk characteristics.

[0091] Module M2: Performs Re-ID modeling on the inspection vehicles to obtain corresponding Re-ID feature information, and stores the Re-ID feature information and parameters in the cloud. The AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model. The vehicle Re-ID model is used to associate vehicle information from the vehicle side and the drone side. The state discrimination model is used to determine vehicle driving anomalies, which are related to sanitation events, including vehicle shaking and inadequate vehicle sealing. The trajectory prediction model is used to identify the consistent trajectory of the vehicle by associating information from multiple sensors.

[0092] Specifically, module M2 includes: Module M2.1: Reading vehicle images from the sanitation cloud business image acquisition library and preprocessing the vehicle images. Module M2.2: Annotating key points in the preprocessed vehicle images. Module M2.3: Dividing the current vehicle into four sides—front, back, left, and right—by locating the key points and using a region segmentation method. Module M2.4: Generating corresponding region masks based on the four sides and associating each region mask with a key point, ensuring that at least one pair of associations exists and they do not overlap. Module M2.5: Extracting region feature vectors from multiple region segmentation results using a CNN convolutional neural network. Module M2.6: Fusing the region feature vectors with global feature vectors. Module M2.7: Performing vehicle re-identification and retrieval on the fused feature vectors to obtain the corresponding feature information.

[0093] Module M3: Constructs a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals.

[0094] Module M4: Collects real-time traffic data, performs preliminary cleaning and processing on the real-time traffic data, and constructs an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model.

[0095] Module M5: Monitors sanitation incidents using the sanitation incident prediction model to obtain a ranking of sanitation risk areas. Module M5 includes: Step S5 involves matching real-time vehicle status data monitored in the selected monitoring area with collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the matched comprehensive data. The calculation formula is as follows:

[0096]

[0097] Among them, a i This represents the current risk value of the vehicle, m. t P represents the weight percentage of road segments in the current time period. tThis represents the risk prediction threshold set for the current time period, and f(t+Δt) represents the subsequent predicted risk coefficient for the current area.

[0098] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0099] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0100] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for monitoring sanitation incidents based on multi-source data fusion and model cascading, characterized in that, include: Step S1: Collect real-time operating data of inspection vehicles and use AI to identify sanitation risk vehicles to obtain corresponding parameters; Step S2: Perform Re-ID modeling on the inspection vehicle to obtain the corresponding Re-ID feature information, and store the Re-ID feature information and the parameters in the cloud; Step S3: Construct a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals. Step S4: Collect real-time traffic data, perform preliminary cleaning and processing on the real-time traffic data, and simultaneously construct an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model. Step S5: Monitor sanitation events using the sanitation event prediction model to obtain a ranking of sanitation risk areas; The AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model; The vehicle Re-ID model is used to associate vehicle information on the vehicle side and the drone side; The state discrimination model is used to determine abnormal vehicle operation, which is related to sanitation incidents, including vehicle shaking and inadequate vehicle sealing. The trajectory prediction model is used to identify the consistent trajectory of a vehicle by associating information from multiple sensors.

2. The sanitation incident monitoring method based on multi-source data fusion and model cascading as described in claim 1, characterized in that, The parameters include real-time vehicle status parameters, high-risk vehicles, and static parameters of the high-risk vehicles. The real-time vehicle status parameters include vehicle speed, driving trajectory, driving lane, driving direction, and driving posture; High-risk vehicles refer to vehicles that could cause sanitation incidents, and are identified through a pre-defined whitelist. The static parameters include vehicle type, license plate, and sanitation risk characteristics.

3. The sanitation incident monitoring method based on multi-source data fusion and model cascading as described in claim 1, characterized in that, Step S2 includes: Step S2.1: Read vehicle images from the sanitation cloud business image acquisition library and preprocess the vehicle images; Step S2.2: Annotate the key points in the preprocessed vehicle image; Step S2.3: By locating the key points and using the region segmentation method, the current vehicle is divided into four sides: front, back, left, and right. Step S2.4: Generate corresponding region masks based on the four faces, and associate each region mask with the key point, wherein there is at least one pair of associations and they do not overlap. Step S2.5: Extract region feature vectors from multiple region segmentation results using a CNN convolutional neural network; Step S2.6: Fuse the region feature vector with the global feature vector; Step S2.7: Perform vehicle re-identification and retrieval on the fused feature vector to obtain the corresponding feature information.

4. The sanitation incident monitoring method based on multi-source data fusion and model cascading as described in claim 1, characterized in that, Step S5 includes matching the real-time vehicle status data monitored in the selected monitoring area with the collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the comprehensive data after matching. The calculation formula is as follows: in, This indicates the vehicle's current risk value. This indicates the weight percentage of road segments in the current time period. This indicates the risk assessment threshold set for the current time period. This indicates the risk coefficient for the current region in future forecasts.

5. A sanitation incident monitoring system based on multi-source data fusion and model cascading, characterized in that, include: Module M1: Collects real-time operating data of inspection vehicles and uses AI to identify sanitation risk vehicles to obtain corresponding parameters; Module M2: Performs Re-ID modeling on the inspection vehicle to obtain the corresponding Re-ID feature information, and stores the Re-ID feature information and the parameters in the cloud; Module M3: Constructs a dynamic topology map of all sanitation risk vehicles by uploading the feature information and parameters to the cloud at fixed time intervals. Module M4: Collects real-time traffic data, performs preliminary cleaning and processing on the real-time traffic data, and constructs an AI multi-model cascade based on the vehicle dynamic topology map to generate a sanitation event prediction model; Module M5: Monitors sanitation events using the sanitation event prediction model to obtain a ranking of sanitation risk areas; The AI ​​multi-model cascade includes a vehicle Re-ID model, a state discrimination model, and a trajectory prediction model; The vehicle Re-ID model is used to associate vehicle information on the vehicle side and the drone side; The state discrimination model is used to determine abnormal vehicle operation, which is related to sanitation incidents, including vehicle shaking and inadequate vehicle sealing. The trajectory prediction model is used to identify the consistent trajectory of a vehicle by associating information from multiple sensors.

6. The sanitation incident monitoring system based on multi-source data fusion and model cascading as described in claim 5, characterized in that, The parameters include real-time vehicle status parameters, high-risk vehicles, and static parameters of the high-risk vehicles. The real-time vehicle status parameters include vehicle speed, driving trajectory, driving lane, driving direction, and driving posture; High-risk vehicles refer to vehicles that could cause sanitation incidents, and are identified through a pre-defined whitelist. The static parameters include vehicle type, license plate, and sanitation risk characteristics.

7. The sanitation incident monitoring system based on multi-source data fusion and model cascading as described in claim 5, characterized in that, Module M2 includes: Module M2.1: Reads vehicle images from the sanitation cloud business image acquisition library and preprocesses the vehicle images; Module M2.2: Annotate key points in the preprocessed vehicle image; Module M2.3: By locating the key points and using the region segmentation method, the current vehicle is divided into four sides: front, back, left, and right. Module M2.4: Generate corresponding region masks based on the four faces, and associate each region mask with the key point, wherein there is at least one pair of such associations and they do not overlap. Module M2.5: Uses a CNN convolutional neural network to extract region feature vectors from multiple region segmentation results; Module M2.6: Merges the regional feature vector with the global feature vector; Module M2.7: Performs vehicle re-identification and retrieval on the fused feature vector to obtain the corresponding feature information.

8. The sanitation incident monitoring system based on multi-source data fusion and model cascading as described in claim 5, characterized in that, The module M5 includes: matching the real-time vehicle status data monitored in the selected monitoring area with the collected static parameters of high-risk vehicles, and calculating the risk value of the current area based on the matched comprehensive data. The calculation formula is as follows: in, This indicates the vehicle's current risk value. This indicates the weight percentage of road segments in the current time period. This indicates the risk assessment threshold set for the current time period. This indicates the risk coefficient for the current region in future forecasts.

Images