Target detection optimization method and system based on multi-source data fusion

Abstract

The application provides a target detection optimization method and system based on multi-source data fusion, relates to the road detection technical field, obtains a first roadbed inclination detection result through an electronic inclinometer; obtains roadbed point cloud data through a laser scanning device, and optimizes and corrects the first roadbed inclination detection result to obtain a second roadbed inclination detection result; vehicle recognition, vehicle roll identification and vehicle speed identification are performed through an image sequence to obtain a plurality of corresponding identification results; then, the vehicle roll information is optimized and corrected; the second roadbed inclination detection result is optimized and corrected according to the correction result to obtain a third roadbed inclination detection result, the technical problems of insufficient accuracy, poor real-time performance and information loss caused by single data in the roadbed inclination detection are solved, the high-precision and real-time roadbed inclination comprehensive detection and analysis are achieved, and then the reliability of road maintenance is ensured, and the road safety is guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road detection, in particular to a target detection optimization method and system based on multi-source data fusion. BACKGROUND

[0002] In the field of road traffic engineering, the inclination of the roadbed directly relates to the safety and service life of the road. The traditional roadbed inclination detection mainly relies on manual measurement or single device data acquisition, which has problems such as low efficiency, limited accuracy, and susceptibility to environmental interference. In addition, with the development of intelligent transportation systems, vehicle driving state data has become a new data source for evaluating road quality, but how to effectively fuse multi-source data to improve the accuracy and real-time performance of roadbed inclination detection has become a technical problem to be solved.

[0003] In summary, the existing technology has the technical problems of insufficient accuracy, poor real-time performance, and information loss caused by single data in roadbed inclination detection. SUMMARY

[0004] The present application provides a target detection optimization method and system based on multi-source data fusion, which is used to solve the technical problems of insufficient accuracy, poor real-time performance, and information loss caused by single data in roadbed inclination detection in the prior art.

[0005] In view of the above problems, the present application provides a target detection optimization method and system based on multi-source data fusion.

[0006] In a first aspect, the present application provides a target detection optimization method based on multi-source data fusion, which comprises:

[0007] Collecting the first roadbed inclination detection result of the target road area roadbed obtained by the electronic inclinometer detection through the Internet of Things; collecting the roadbed point cloud data of the target road area roadbed obtained by scanning the laser scanning device through the Internet of Things, performing roadbed fitting to obtain the roadbed inclination fitting result, and optimizing and correcting the first roadbed inclination detection result to obtain the second roadbed inclination detection result; collecting the passing image sequence of the target road area when multiple vehicles pass through the Internet of Things, performing vehicle recognition, vehicle roll identification and vehicle speed identification to obtain multiple vehicle feature information, multiple vehicle roll information and multiple vehicle speed information; according to the multiple vehicle speed information combined with the multiple vehicle feature information, the multiple vehicle roll information is optimized and corrected to obtain multiple corrected vehicle roll information; according to the multiple corrected vehicle roll information, multiple roadbed inclination analysis results are obtained by classification, and the second roadbed inclination detection result is optimized and corrected to obtain the third roadbed inclination detection result.

[0008] In a second aspect, the application provides a target detection optimization system based on multi-source data fusion, comprising:

[0009] The roadbed data acquisition module is configured to acquire first roadbed tilt detection results of a roadbed of a target road region detected by an electronic inclinometer through the Internet of Things.

[0010] The one or more technical solutions provided in the application have at least the following technical effects or advantages:

[0011] The target detection optimization method based on multi-source data fusion provided by the application acquires first roadbed tilt detection results of a roadbed of a target road region detected by an electronic inclinometer through the Internet of Things. The method acquires roadbed point cloud data of the roadbed of the target road region scanned by a laser scanning device through the Internet of Things, performs roadbed fitting, acquires roadbed tilt fitting results, and optimizes and corrects the first roadbed tilt detection results to acquire second roadbed tilt detection results. The method acquires a passing image sequence of multiple vehicles passing through the target road region through the Internet of Things, performs vehicle identification, vehicle roll identification, and vehicle speed identification, and acquires multiple vehicle feature information, multiple vehicle roll information, and multiple vehicle speed information. The method optimizes and corrects the multiple vehicle roll information according to the multiple vehicle speed information in combination with the multiple vehicle feature information to acquire multiple corrected vehicle roll information. The method classifies and acquires multiple roadbed tilt analysis results according to the multiple corrected vehicle roll information, optimizes and corrects the second roadbed tilt detection results, and acquires third roadbed tilt detection results. The method solves the technical problems of insufficient accuracy, poor real-time performance, and information loss caused by single data in the roadbed tilt detection in the prior art. Through multi-source data fusion, the method improves the accuracy and real-time performance in roadbed tilt detection, achieves high-precision and real-time roadbed tilt comprehensive detection and analysis, and further ensures the reliability of road maintenance and road safety. Attached Figure Description

[0012] Figure 1 This application provides a schematic diagram of the target detection optimization method based on multi-source data fusion.

[0013] Figure 2 This application provides a schematic diagram of the target detection optimization system structure based on multi-source data fusion.

[0014] Explanation of reference numerals in the attached diagram: 11 Roadbed data acquisition module, 12 Roadbed tilt optimization and correction module, 13 Vehicle image recognition module, 14 Vehicle side tilt optimization and correction module, 15 Roadbed tilt detection result acquisition module. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0016] Example 1, as Figure 1 As shown, this application provides a target detection optimization method based on multi-source data fusion, the method comprising:

[0017] Step S100: Collect the first roadbed tilt detection results of the target road area obtained by electronic inclinometer detection through the Internet of Things.

[0018] Specifically, the detection optimization target in this embodiment mainly focuses on roadbed detection. In this scenario, even a slight tilt of the roadbed can pose varying degrees of safety hazards to vehicle traffic, a danger particularly pronounced at curves. Because vehicles traveling at high speeds through curves generate strong centrifugal forces, these forces can gradually exacerbate the risk of roadside collapse, seriously threatening driving safety. Therefore, to ensure the stability of the road structure and driving safety, a comprehensive detection optimization scheme is proposed. This scheme aims to achieve comprehensive and accurate monitoring and real-time early warning of roadbed tilt by integrating high-precision radar point cloud data, vehicle dynamic tilt analysis, and advanced data processing technologies. This scheme not only enables timely detection and location of the specific position and degree of roadbed tilt but also allows for dynamic optimization and correction of the monitoring results based on actual feedback during vehicle travel, thereby ensuring the timeliness and effectiveness of road maintenance.

[0019] In the construction of the detection optimization system for the subgrade inclination of the target road area, first of all, the efficient integration and remote monitoring capability of the Internet of Things technology is used to realize intelligent monitoring of the subgrade inclination state. The Internet of Things realizes real-time sensing, transmission and analysis of various physical parameters through sensors, communication networks and intelligent processing platforms. In this embodiment, the Internet of Things is used to integrate detection devices such as electronic inclinometers to realize automatic collection and remote monitoring of subgrade inclination data. The electronic inclinometer is a high-precision measuring device that senses the inclination angle through the built-in sensor and converts it into an electrical signal for output. It is usually installed at key positions of the road subgrade, such as the inside and outside of curves, near bridge supports, etc., to monitor the inclination of these areas.

[0020] The electronic inclinometer is connected to the Internet of Things system and establishes a communication connection with the data acquisition terminal. The electronic inclinometer continuously senses the inclination angle of the subgrade and transmits the measurement data to the data acquisition terminal in real time. The data acquisition terminal uploads the collected inclination data to the data center through the Internet of Things network for subsequent processing. Further, the data center calculates the inclination angle, inclination direction and other key parameters of the subgrade according to the measurement data of the electronic inclinometer, and arranges the first subgrade inclination detection result according to these parameters, including inclination position, inclination degree, timestamp and other information. For example, assuming that the electronic inclinometer deployed on the inside of a curve continuously monitors the subgrade inclination angle from 0.5° to 1.2°, the system will trigger the alarm mechanism when the preset threshold (such as 1°) is reached, and generate the first subgrade inclination detection result. The result records the position of the inclination (such as "curve inside X section"), the inclination degree (such as "1.2°"), the monitoring time (such as "2023-XX-XX XX:XX"), and other information, providing an important basis for subsequent optimization and correction work.

[0021] By using the Internet of Things technology to collect detection data from electronic inclinometers, and generating the first subgrade inclination detection result of the target road area, a solid foundation is laid for subsequent optimization and correction combined with radar point cloud fitting, vehicle roll analysis and other multi-source data.

[0022] Step S200: Collect the subgrade point cloud data of the subgrade of the target road area scanned by the laser scanning device through the Internet of Things, perform subgrade fitting to obtain the subgrade inclination fitting result, optimize and correct the first subgrade inclination detection result, and obtain the second subgrade inclination detection result.

[0023] Further, a laser scanning device is introduced to optimize and correct the preliminary roadbed inclination detection results. The laser scanning device is a high-precision, non-contact three-dimensional measurement device. By emitting a laser beam and receiving its reflected light, it can quickly obtain the three-dimensional coordinate information of the target surface and generate dense point cloud data. In the target road area, especially in key parts such as curves and bridges, laser scanning devices are deployed. The devices are connected to the data collection system through Internet of Things technology, enabling remote control and real-time data transmission. After starting, the laser scanning device scans the roadbed comprehensively and accurately. The device can automatically adjust the scanning parameters to ensure that the obtained point cloud data is both comprehensive and detailed. After scanning is completed, the laser scanning device transmits the generated massive point cloud data to the data center through the Internet of Things network for subsequent processing and analysis.

[0024] The data center performs preprocessing operations such as denoising, filtering, and thinning on the received point cloud data to improve data quality and reduce computational load. In addition, point cloud data from different perspectives under multiple laser scanning devices are registered to ensure they are in the same coordinate system. Further, feature information such as roadbed edges and slope changes is extracted from the preprocessed point cloud data, and a three-dimensional geometric model of the roadbed is fitted based on this feature information. That is, different inclination roadbed anchor frames are designed according to the geometric characteristics of the roadbed. These anchor frames are pre-set geometric frames with different inclination angles, used to simulate possible roadbed inclination shapes. Next, the least squares method or other optimization algorithms are used to fit the roadbed point cloud data with each roadbed anchor frame, and the anchor frame with the highest fitting degree is selected by calculating the matching degree (such as mean square error, fitting index, etc.) of each anchor frame and the point cloud data. The inclination angle corresponding to this anchor frame is the roadbed inclination angle obtained by analyzing the laser scanning data, which more accurately and comprehensively reflects the actual shape of the roadbed.

[0025] Finally, the inclination angle in the first roadbed inclination detection result obtained by the electronic inclinometer is fused with the inclination angle obtained by laser scanning analysis. The fusion process can use a two-weighted calculation method, that is, different weights are assigned to the two detection methods according to their accuracy, reliability, and actual application scenario requirements. For example, if the accuracy of the laser scanning data is significantly higher than that of the electronic inclinometer, the laser scanning result can be given a higher weight. Through weighted averaging, the second roadbed inclination detection result is obtained. The second roadbed inclination detection result not only integrates the experience of traditional detection methods, but also fully utilizes the high-precision and high-efficiency advantages of Internet of Things technology and laser scanning technology, achieving comprehensive and accurate detection of roadbed inclination. In practical applications, this scheme helps to timely detect roadbed diseases and prevent road safety accidents, providing scientific decision-making basis for road maintenance and management.

[0026] For example, assume that in the detection of a roadbed of a certain highway, the electronic inclinometer initially measures an inclination of 2.5°, and then the laser scanning device collects a large amount of roadbed point cloud data, and finds that the anchor frame with an inclination of 2.7° has the highest matching degree with the point cloud data through fitting with multiple anchor frames with different inclinations. In the two-weighted calculation, if the laser scanning result is given a weight of 0.7 and the electronic inclinometer result is given a weight of 0.3, then the final calculated second roadbed inclination detection result is (2.5°×0.3+2.7°×0.7)=2.64°, which is closer to the true inclination state of the roadbed.

[0027] By two-weighted optimization correction of the first roadbed inclination detection result after matching of the laser point cloud data, the actual road surface situation can be more truly reflected, and the accuracy and efficiency of the roadbed inclination detection are improved.

[0028] Step S300: Collect the passing image sequences of multiple vehicles passing through the target road area through the Internet of Things, perform vehicle recognition, vehicle roll identification, and vehicle speed identification, and obtain multiple vehicle feature information, multiple vehicle roll information, and multiple vehicle speed information.

[0029] Optionally, multiple high-definition cameras are installed at the target road area, especially at the curve, which have night vision, wide-angle, and high-speed continuous shooting functions, and can record detailed images of passing vehicles all day long and without dead angles. As the vehicles enter and exit the curve, the cameras will continuously shoot and generate a series of high-definition images, forming image sequences when the vehicles pass through. These image sequences not only contain the appearance features of the vehicles, but also imply the motion state information of the vehicles. After collecting and analyzing these image sequences through the Internet of Things, multiple vehicles are identified and processed, laying a foundation for subsequent analysis.

[0030] The collected image sequences are analyzed frame by frame using advanced image recognition and deep learning algorithms. First, the vehicle main body in the image is identified, and then the brand, model, and possible weight level of the vehicle are accurately identified by comparing the vehicle appearance features (such as car light shape, body lines, brand logo, etc.) with the preset vehicle database, completing multiple vehicle identification and obtaining multiple vehicle feature information. This process relies on a large vehicle feature database and a continuously optimized learning model to ensure the accuracy and efficiency of the identification.

[0031] Vehicle roll is an important indicator for evaluating the stability of vehicles driving on curves. Based on image recognition, further analysis of the changes in the vehicle's profile, particularly the inclination angle of the vehicle body relative to the horizontal plane, can be performed. By calculating the pixel-level displacement of the vehicle in the image and combining the geometric correction parameters of the camera, the roll angle of the vehicle can be accurately calculated, and multiple vehicle roll information can be obtained. For vehicles with a roll angle exceeding the pre-set safety threshold, the system will immediately issue a warning to remind the driver to pay attention to road safety or take appropriate measures. For example, a heavy truck is driving at high speed on a curve, and its roll angle gradually increases due to centrifugal force. The camera deployed on the outside of the curve captures this change and transmits it in real time to the image recognition system. The system quickly identifies the brand and model of the truck and calculates that its roll angle has exceeded the safety range by analyzing the changes in the truck's profile in the image. At this time, the system will automatically trigger the warning mechanism and send warning information to the truck driver, and may also synchronize relevant information to the traffic management center for timely further measures.

[0032] At the same time, in order to comprehensively evaluate the driving state of the vehicle, the system will also use the timestamp information in the image sequence, combined with the pixel displacement of the vehicle, to calculate the change in the position of the vehicle between adjacent frames to estimate the real-time speed of the vehicle, obtaining multiple vehicle speed information. This technique is not only applicable to straight roads but also to complex road conditions such as curves, providing more accurate vehicle speed monitoring means for traffic management departments.

[0033] Through the Internet of Things technology to collect image sequences of target road areas, and combined with advanced image recognition and deep learning algorithms, comprehensive real-time identification and monitoring of vehicle types, roll angles, and speeds are achieved. This process not only obtains multiple real-time data sources but also improves the timeliness of target detection and the intelligence level of traffic management, providing strong support for ensuring road traffic safety and optimizing road design.

[0034] Step S400: According to the multiple vehicle speed information combined with the multiple vehicle characteristic information, the multiple vehicle roll information is optimized and corrected to obtain multiple corrected vehicle roll information.

[0035] For example, in-depth analysis of the behavior characteristics of vehicles driving on curves, combined with multiple vehicle speed information and vehicle characteristic information collected by the Internet of Things, a process of optimizing and correcting vehicle roll information can be performed. The core of this process is to identify and eliminate the roll caused by the dynamic characteristics of the vehicle itself (such as speed and weight), so as to more accurately evaluate the roll impact on the vehicle caused by the roadbed inclination. This method is also applicable to non-curve areas.

[0036] Specifically, multiple vehicle speed information, vehicle characteristic information, and original vehicle roll information from the Internet of Things system are collected. The vehicle speed information records the actual speed of each vehicle when passing through the curve; the vehicle characteristic information contains key parameters such as the brand, model, and possible weight of the vehicle; and the vehicle roll information is directly obtained through image recognition technology, which contains the roll angle of the vehicle during driving due to various reasons, including roadbed inclination and vehicle dynamics. Based on physical principles and vehicle dynamics knowledge, a vehicle turning roll prediction model is constructed, which considers the influence of vehicle speed, weight, and curve curvature on roll. Generally, the greater the vehicle speed, the greater the centrifugal force generated during turning, resulting in greater roll; at the same time, heavy vehicles will also produce greater roll during turning due to their large inertia. In addition, the curvature radius of the curve also affects the degree of roll, and the smaller the curvature, the more obvious the roll. Then, using the above model and vehicle characteristic information (such as weight) and speed information, the possible self-roll of each vehicle during driving on the curve is predicted, and the self-roll caused by vehicle dynamics is separated from the total roll information. Finally, the predicted vehicle self-turning roll angle is subtracted from the original vehicle roll information, which realizes the optimization correction of the roll information, making the remaining part more accurately reflect the vehicle roll caused by roadbed inclination, and ensuring that the evaluation of the influence of roadbed inclination is not disturbed by the vehicle's own dynamic characteristics.

[0037] The multiple corrected vehicle roll information is output and used for subsequent analysis, which provides an important basis for evaluating roadbed inclination, judging road safety, and formulating maintenance measures. For example, if it is found that a certain section of road shows a high corrected roll degree when multiple vehicles pass through, it may mean that the roadbed of this section of road has a serious inclination problem and needs to be repaired or reinforced immediately.

[0038] Step S500: According to the multiple corrected vehicle roll information, multiple roadbed inclination analysis results are obtained by classification, and the second roadbed inclination detection result is optimized and corrected to obtain a third roadbed inclination detection result.

[0039] Specifically, the multiple corrected vehicle roll information is classified according to vehicle type, speed interval, passing time, etc. The purpose of classification is to identify the influence mode of roadbed inclination on vehicle roll under different conditions, so as to more accurately evaluate the inclination of the roadbed. Then, based on the results of classification analysis, statistical methods and data mining techniques are used to generate multiple roadbed inclination analysis results, which include the average value, maximum value, and change trend of roadbed inclination in different positions and time periods, and can fully reflect the roadbed inclination.

[0040] Next, the second subgrade inclination detection result (i.e., the result obtained by synthesizing the electronic inclinometer and the laser scanning data before) is fused with the subgrade inclination analysis result. Various factors can be considered during the fusion process, such as the reliability of the correction vehicle roll information, the accuracy of the second subgrade inclination detection result, the statistical significance of the classification analysis, etc. The second subgrade inclination detection result is further optimized and corrected through weighted averaging, Bayesian fusion or other advanced fusion algorithms, thereby eliminating or reducing the errors that may exist in a single detection method and improving the accuracy and reliability of the final detection result. For example, on a certain curved road, it is found through classification analysis that the correction roll information of heavy trucks when passing at high speed (such as speed exceeding 60 km / h) is generally high and shows a gradually increasing trend over time. At the same time, the second subgrade inclination detection result shows that there is about 2% inclination on this section of road. Combining these information, it can be considered that the subgrade inclination of this section of road may be greater than 2% and there is a risk of further deterioration. Therefore, in the optimization and correction process, the correction roll information of heavy trucks when passing at high speed can be given a higher weight, so as to obtain a third subgrade inclination detection result that is closer to the actual situation.

[0041] Finally, the third subgrade inclination detection result after optimization and correction is output and used for practical application. Through the optimization and constraint of multi-dimensional data, the accuracy of subgrade inclination detection is improved. Moreover, these results not only provide a scientific basis for road maintenance departments to help them develop targeted maintenance plans, but also provide important safety warning information for traffic management departments to help them take timely measures to ensure road traffic safety.

[0042] Further, the first subgrade inclination detection result of the subgrade of the target road region obtained by the electronic inclinometer detection through the Internet of Things, the step S100 of the present application further comprises:

[0043] Step S110: configuring an electronic inclinometer at the subgrade detection position of the target road region, and connecting based on the Internet of Things, wherein the target road region is a curved road region.

[0044] Step S120: collecting a plurality of subgrade inclination results of the subgrade obtained by the electronic inclinometer test through the Internet of Things, and calculating a first subgrade inclination detection result, wherein each subgrade inclination result comprises a subgrade inclination.

[0045] In this embodiment, the preferred target road area is the curve area. The specific monitoring target is the roadbed inclination of a certain road area. Because the terrain of the curve area is complex and the stress changes greatly during vehicle driving, roadbed deformation is easy to occur, which poses a potential threat to driving safety. Therefore, the curve area is selected as the monitoring target. A high-precision electronic inclinometer is selected as the monitoring tool. This device can accurately measure the inclination angle of an object (such as a roadbed) in different directions and is an important means of evaluating roadbed stability. Next, representative roadbed detection positions are selected in the curve section of the target road area for the installation of electronic inclinometers. These positions should cover different curvature sections of the curve to comprehensively reflect the overall inclination of the roadbed. During installation, it is necessary to ensure that the inclinometer is stable and in close contact with the roadbed to reduce measurement errors. The electronic inclinometer is connected to a remote monitoring center or data processing system using Internet of Things technology to realize real-time data transmission and remote monitoring. The connection of the Internet of Things ensures the immediacy and accuracy of the monitoring data, providing a solid foundation for subsequent data analysis.

[0046] Through the Internet of Things platform, roadbed inclination data obtained by electronic inclinometer testing is continuously and automatically collected. These data include the inclination of the roadbed at different time points, reflecting the dynamic change process of the roadbed. Each roadbed inclination result is presented in digital form, including but not limited to inclination angle (in degrees or radians), measurement timestamp, and other key information. These data will serve as the basis for subsequent analysis. The collected multiple roadbed inclination results are analyzed comprehensively, and the first roadbed inclination detection result is calculated through statistical methods (such as mean, standard deviation, trend line, etc.). This step aims to extract the overall characteristics of roadbed inclination, such as average inclination and inclination change trend, providing important evidence for subsequent roadbed stability evaluation. Further, based on the first roadbed inclination detection result, the current stability state of the roadbed in the curve area is evaluated. If the inclination exceeds the preset threshold or shows a significant growth trend, it may indicate that the roadbed has potential safety hazards, and further measures need to be taken for reinforcement or repair. The monitoring results are fed back to the road management department or maintenance team in a timely manner to provide data support for developing targeted maintenance plans. For example, road segments with high inclination can be prioritized for reinforcement projects, and areas with a clear inclination trend can have increased daily patrols and monitoring frequency.

[0047] Furthermore, through the Internet of Things, roadbed point cloud data of the roadbed of the target road area scanned by the laser scanning device is collected, roadbed fitting is performed, the roadbed inclination fitting result is obtained, and the first roadbed inclination detection result is optimized and corrected. Step S200 of the present application further includes:

[0048] Step S210: Collecting roadbed point cloud data of the roadbed of the target road scanned by the laser scanning device through the Internet of Things.

[0049] Step S220: Obtain a plurality of sample roadbed fitting anchor frames, wherein the roadbed inclinations in the plurality of sample roadbed fitting anchor frames are different.

[0050] Step S230: Traverse fitting in the roadbed point cloud data using the plurality of sample roadbed fitting anchor frames to obtain a sample roadbed fitting anchor frame with the largest roadbed fitting degree as a roadbed inclination fitting result.

[0051] Step S240: Output the roadbed inclination of the sample roadbed fitting anchor frame in the roadbed inclination fitting result, and perform optimization correction calculation on the first roadbed inclination detection result to obtain a second roadbed inclination detection result.

[0052] Optionally, a laser scanning device integrated with Internet of Things technology is used to perform high-precision scanning of the roadbed of the target road area. During this process, the laser scanning device emits a laser beam and receives the signal reflected back from the road surface. By calculating the round-trip time difference of the laser and the angle information, detailed roadbed point cloud data is generated. These point cloud data are a collection of points in three-dimensional space, accurately recording the shape and height changes of the roadbed surface, providing a basis for subsequent roadbed inclination analysis. In the data processing stage, a plurality of sample roadbed fitting anchor frames can be prepared in advance. These anchor frames are theoretical models designed based on different roadbed inclinations, representing possible roadbed inclination forms. Each anchor frame contains a series of preset inclination parameters, such as inclination angle and inclination direction, for matching and fitting with actual point cloud data. Next, an algorithm is used to automatically traverse these sample roadbed fitting anchor frames in the collected roadbed point cloud data. During the traversal process, the algorithm attempts to match each anchor frame with the point cloud data, and evaluates the fitting effect by calculating the fitting degree (such as mean square error, correlation coefficient, etc.) of the anchor frame and the point cloud data. After the traversal is completed, the sample roadbed fitting anchor frame with the highest fitting degree is selected. This anchor frame is the best fitting result of the roadbed inclination, reflecting the actual inclination of the target road roadbed. After obtaining the roadbed inclination fitting result, the first roadbed inclination detection result obtained initially is optimized and corrected according to the roadbed inclination information in the result. This step involves fine-tuning the inclination angle, eliminating errors, or re-evaluating the detection result. Finally, the second roadbed inclination detection result after optimization and correction is output. This result is not only more accurate and reliable, but also provides an important reference for subsequent work such as road maintenance and design optimization.

[0053] Through the above steps, not only the precise detection of the roadbed inclination is realized, but also the detection efficiency and accuracy are improved through the integration of Internet of Things and laser scanning technology, bringing significant technological progress to the field of road engineering.

[0054] Further, the multiple sample roadbed fitting anchor frames are used to traverse fitting in the roadbed point cloud data, and a sample roadbed fitting anchor frame with the maximum roadbed fitting degree is obtained. The step S230 of the present application further includes:

[0055] Step S231: A first sample roadbed fitting anchor frame is randomly selected from the multiple sample roadbed fitting anchor frames.

[0056] Step S232: The first sample roadbed fitting anchor frame is used to perform iterative fitting in the roadbed point cloud data. In each fitting, the proportion of the point cloud falling into the first sample roadbed fitting anchor frame in the roadbed point cloud data is calculated as the roadbed fitting degree. Finally, the number of convergent iterations is reached, and the maximum first roadbed fitting degree is output.

[0057] Step S233: The sample roadbed fitting anchor frame is continuously selected from the multiple sample roadbed fitting anchor frames to perform iterative fitting in the roadbed point cloud data, and multiple roadbed fitting degrees are obtained.

[0058] Step S234: The sample roadbed fitting anchor frame with the maximum roadbed fitting degree is output as the roadbed inclination fitting result.

[0059] Further, one is randomly selected from the multiple sample roadbed fitting anchor frames as the first sample roadbed fitting anchor frame, so as to reduce the initial dependence of the algorithm and increase the flexibility of fitting. Subsequently, the first sample roadbed fitting anchor frame is used to perform iterative fitting in the roadbed point cloud data (i.e., a three-dimensional point set of the roadbed surface obtained through laser scanning, photogrammetry, etc.). During the iteration, the position, size or inclination angle of the anchor frame is adjusted according to the distribution of the point cloud data to maximize the matching degree of the anchor frame and the point cloud data. After each iteration, the proportion of the number of point clouds falling into the current anchor frame in the entire roadbed point cloud data is calculated, and this proportion is defined as the roadbed fitting degree, which directly reflects the similarity between the current anchor frame and the actual roadbed form. When the number of iterations reaches the preset convergent fitting number (i.e., the number threshold at which the algorithm considers that the fitting degree will no longer be significantly improved by further iteration), the iteration is stopped, and the maximum first roadbed fitting degree and the corresponding anchor frame state calculated at this time are output.

[0060] After completing the first round of random selection and fitting, the remaining sample roadbed fitting anchor frames are continuously traversed, and the iterative fitting process is repeated. Independent iterative fitting is performed for each traversed anchor frame, and the corresponding roadbed fitting degree is calculated. In this way, multiple roadbed fitting degrees under different anchor frames are obtained. The maximum roadbed fitting degree and the corresponding sample roadbed fitting anchor frame are found among all the calculated roadbed fitting degrees. This anchor frame is considered to be the result that best represents the actual roadbed inclination state, i.e., the roadbed inclination fitting result. In summary, the accurate fitting result of the roadbed inclination is obtained by fitting and selection, which improves the accuracy of roadbed inclination detection.

[0061] Further, the target road area is covered by an Internet of Things system, which includes high-definition cameras, sensor networks, and other devices to capture image sequences and other related data in real time. At the same time, a sample image set and a corresponding sample vehicle set are prepared for training the vehicle recognizer. These sample data should include a variety of vehicle images under different angles, lighting conditions, and vehicle types to ensure the generalization ability of the recognizer. The sample image set refers to a set of pre-collected and labeled images that show vehicles passing through a specific area (such as a road or intersection) under different conditions (such as different times, weather, and lighting). These images are usually sourced from actual monitoring cameras, simulation datasets, or public dataset resources. They contain rich information such as vehicle appearance, size, color, driving direction, lane position, and background environment such as road signs, other vehicles, pedestrians, and buildings. To train an efficient and accurate vehicle recognition model, the sample image set needs to cover as many vehicle types and complex scenarios as possible to ensure the model's generalization ability. The sample vehicle set is a set of data corresponding to the sample image set, which contains the annotation information of each vehicle in the sample image. Specifically, for each image in the sample image set, the sample vehicle set provides accurate recognition results of all vehicles in the image, including but not limited to vehicle type (such as sedan, truck, motorcycle, etc.), brand, color, license plate number (if visible and allowed to be identified), position coordinates (specific position in the image), and other information.

[0062] Step S310: Vehicle recognition is performed on the multiple passing images in the multiple passing image sequences when multiple vehicles pass through, to obtain a multiple identified vehicle set. The vehicle recognizer is trained by collecting a sample passing image set and a sample identified vehicle set to perform vehicle recognition.

[0063] Step S320: Based on the slowfast network, vehicle side tilt recognition and vehicle speed recognition are performed on the multiple passing image sequences to obtain multiple vehicle side tilt information and multiple vehicle speed information.

[0064] In this embodiment, an Internet of Things system is constructed to cover the target road area, which includes high-definition cameras, sensor networks, and other devices to capture image sequences and other related data in real time. At the same time, a sample image set and a corresponding sample vehicle set are prepared for training the vehicle recognizer. These sample data should include a variety of vehicle images under different angles, lighting conditions, and vehicle types to ensure the generalization ability of the recognizer. The sample image set refers to a set of pre-collected and labeled images that show vehicles passing through a specific area (such as a road or intersection) under different conditions (such as different times, weather, and lighting). These images are usually sourced from actual monitoring cameras, simulation datasets, or public dataset resources. They contain rich information such as vehicle appearance, size, color, driving direction, lane position, and background environment such as road signs, other vehicles, pedestrians, and buildings. To train an efficient and accurate vehicle recognition model, the sample image set needs to cover as many vehicle types and complex scenarios as possible to ensure the model's generalization ability. The sample vehicle set is a set of data corresponding to the sample image set, which contains the annotation information of each vehicle in the sample image. Specifically, for each image in the sample image set, the sample vehicle set provides accurate recognition results of all vehicles in the image, including but not limited to vehicle type (such as sedan, truck, motorcycle, etc.), brand, color, license plate number (if visible and allowed to be identified), position coordinates (specific position in the image), and other information.

[0065] A plurality of passing image sequences are obtained from the Internet of Things system. First, preprocessing is performed, including denoising, contrast enhancement, size adjustment, etc., to improve image quality and facilitate subsequent processing. Further, a high-efficiency vehicle identifier is trained using the preprocessed sample passing image set and the sample vehicle set. This identifier can use deep learning algorithms such as convolutional neural networks, which can accurately identify vehicles in images and distinguish their types through continuous iterative learning. The trained vehicle identifier is applied to the plurality of real-time captured passing image sequences. For each frame of the sequence, the identifier analyzes and identifies the vehicles in the frame, and generates a plurality of vehicle sets, each containing all the vehicle information identified within a specific time period, such as vehicle type, location, etc.

[0066] To capture both rapid changes and slow trends in vehicle motion, a slowfast network is used for vehicle roll and speed identification. The slowfast network is a video processing network that captures multi-level temporal information in videos by processing image sequences at different speeds in parallel. The network consists of two parts: the slow path captures spatial semantic information, while the fast path focuses on capturing temporal details. The plurality of passing image sequences is input into a specific branch of the slowfast network, which analyzes vehicle outlines, wheel positions, and other features, and combines temporal sequence information to determine whether the vehicle is rolling and outputs a plurality of vehicle roll information, including roll angle and direction. Meanwhile, another branch of the slowfast network is used to calculate and output a plurality of vehicle speed information by combining the displacement changes between consecutive frames, which involves accurate tracking of vehicle positions and accurate measurement of inter-frame time differences.

[0067] For example, on a highway, the Internet of Things system captures a series of passing vehicle image sequences. First, the system uses the trained vehicle identifier to identify a plurality of different types of cars, trucks, etc. from the sequence and records their positions and types. Then, the slowfast network analyzes the image sequences of these vehicles in depth and finds that one of the trucks has a slight roll when turning, so it outputs the roll angle and direction information of the truck. At the same time, the network calculates the real-time speed of each vehicle according to the displacement changes between different frames and provides the corresponding speed information. Through such processing, the Internet of Things system can not only capture real-time vehicle information on the road, but also accurately identify the roll state and speed of the vehicle, providing strong support for road safety monitoring and traffic management.

[0068] Further, based on the slowfast network, the vehicle roll identification and vehicle speed identification of the plurality of passing image sequences are performed, and the step S320 of the present application further includes:

[0069] Step S321: According to a first extraction step, the multiple through image sequences are respectively subjected to through image extraction and down-sampling processing, and multiple first through image sets are obtained.

[0070] Step S322: According to a second extraction step greater than the first extraction step, the multiple through image sequences are respectively subjected to through image extraction, and multiple second through image sets are obtained.

[0071] Step S323: Based on the historical monitoring data of the target road area, multiple sample first through image sets and multiple sample second through image sets are collected, and a sample vehicle roll information set and a sample vehicle speed information set are obtained.

[0072] Step S324: Using the multiple sample first through image sets, the multiple sample second through image sets, the sample vehicle roll information set, and the sample vehicle speed information set, a vehicle through feature recognizer is trained based on a slowfast network.

[0073] Step S325: Using the vehicle through feature recognizer, the multiple first through image sets and the multiple second through image sets are inputted for recognition, and multiple vehicle roll information and multiple vehicle speed information are obtained.

[0074] Specifically, the multiple continuous image sequences (i.e., continuous frames in a video stream) obtained are preprocessed to adapt to the input requirements of the slowfast network. This step needs to be divided into two sub-steps, namely first extraction step processing and second extraction step processing. The first extraction step processing refers to extracting part of the frames from the original image sequence according to a pre-set first extraction step (e.g., extracting 1 frame every 5 frames), and performing down-sampling processing (e.g., reducing the image resolution from 1080p to 720p) to reduce the amount of calculation while retaining key information. This process generates multiple first through image sets, each containing image sequences extracted by the step and down-sampled. The extraction step refers to the interval of selecting frames from the original image sequence. The down-sampling is a process of reducing the image resolution to reduce the amount of data and computational complexity. Subsequently, a larger extraction step (e.g., extracting 1 frame every 10 frames) is used to extract frames from the original image sequence again, but this time without down-sampling processing. This is done to obtain more sparse but time-span larger image information for the slow path in the slowfast network. This process generates multiple second through image sets.

[0075] Further, based on the historical monitoring data of the target road area, a certain number of sample image sets are collected, including a plurality of sample first passing image sets and a plurality of sample second passing image sets. At the same time, a set of vehicle roll information and a set of vehicle speed information corresponding to these image sets need to be obtained. Using the prepared sample data (including sample image sets and corresponding vehicle roll and speed information), a vehicle passing feature recognizer is trained based on the slowfast network architecture. Since the recognizer is based on the slowfast network, it combines the fast path (processing the first passing image set, capturing details and rapid changes) and the slow path (processing the second passing image set, capturing global and slow changes) to achieve efficient and accurate feature extraction.

[0076] Next, the trained vehicle passing feature recognizer is used to process new image sequences (i.e. real-time or newly collected image sequences) to identify vehicle roll information and speed information in each frame of image by inputting a plurality of first passing image sets and a plurality of second passing image sets. Finally, a plurality of recognition results containing vehicle roll and speed information are output. For example, when a new video stream enters the system, image extraction and downsampling (or only extraction) are first performed according to the preset extraction step, and then the image sets are input into the trained slowfast network. The network outputs the roll angle and speed estimate of the vehicle in each frame of image. The acquisition of this information ensures the accuracy of subsequent roadbed tilt detection.

[0077] Further, according to the plurality of vehicle speed information and the plurality of vehicle feature information, the plurality of vehicle roll information is optimized and corrected to obtain a plurality of corrected vehicle roll information. The step S400 of the present application further comprises:

[0078] Step S410: based on the vehicle roll test data of the same road area of the target road area, a set of sample vehicle feature information, a set of sample vehicle speed information, and a set of sample speed roll information are obtained.

[0079] Step S420: using the set of sample vehicle feature information, the set of sample vehicle speed information, and the set of sample speed roll information, a speed roll analyzer is trained.

[0080] Step S430: using the speed roll analyzer, speed roll analysis is performed on the plurality of vehicle speed information and the plurality of vehicle feature information to obtain a plurality of speed roll information.

[0081] Step S440: reducing the plurality of speed roll information in the plurality of vehicle roll information to obtain a plurality of corrected vehicle roll information.

[0082] For example, a series of vehicle roll test data is collected based on the same road area of the target road area, which includes not only the roll angle of the vehicle under different conditions (i.e., original roll information), but also the corresponding vehicle feature information (such as vehicle type, load, suspension system, etc.) and vehicle speed information. These data form the basis for training the speed roll analyzer. Next, a speed roll analyzer is trained using the collected sample vehicle feature information set, sample vehicle speed information set, and sample speed roll information set. The purpose of this analyzer is to learn the complex relationship between vehicle features, speed, and roll, so as to be able to predict or correct the vehicle roll angle under given speed and features. After training is completed, the speed roll analyzer is applied to new vehicle data. For each set of vehicle speed information and vehicle feature information, the analyzer outputs a predicted or corrected speed roll information, which takes into account the effects of vehicle speed and features on the roll angle, thereby improving the accuracy of the roll information. Finally, the speed roll information obtained by the speed roll analysis is applied to the original vehicle roll information for optimization and correction. Specifically, the corresponding speed roll information can be reduced (or replaced) within multiple vehicle roll information, thereby obtaining multiple corrected vehicle roll information, which is closer to the actual roll of the vehicle during actual driving.

[0083] By using vehicle speed information and vehicle feature information to optimize and correct vehicle roll information, not only the accuracy of the roll information is improved, but also more reliable data support is provided for subsequent traffic monitoring, safety warning, and other applications.

[0084] Further, according to the plurality of corrected vehicle roll information, a plurality of roadbed tilt analysis results are obtained by classification, and the second roadbed tilt detection result is optimized and corrected to obtain a third roadbed tilt detection result. The step S500 of the present application further includes:

[0085] Step S510: According to the monitoring data of vehicle roll and roadbed tilt, a sample corrected vehicle roll information set and a sample roadbed tilt analysis result set are collected, and each sample roadbed tilt analysis result includes a roadbed tilt degree.

[0086] Step S520: A classification mapping relationship of the sample corrected vehicle roll information set and the sample roadbed tilt analysis result set is constructed.

[0087] Step S530: According to the plurality of corrected vehicle roll information, a plurality of roadbed tilt analysis results are obtained by classification.

[0088] Step S540: The plurality of roadbed tilt analysis results are used to optimize and correct the second roadbed tilt detection result to obtain a third roadbed tilt detection result.

[0089] Specifically, according to the monitored vehicle roll and subgrade inclination, a set of sample corrected vehicle roll information and a set of sample subgrade inclination analysis results are obtained, each sample subgrade inclination analysis result including a subgrade inclination degree. Next, in order to accurately infer the subgrade inclination condition from the vehicle roll information, a mathematical model or classifier needs to be constructed, which can learn and map the relationship between the set of sample corrected vehicle roll information and the set of sample subgrade inclination analysis results. Through the training of the model, a classification mapping relationship can be established, so that given a set of vehicle roll information, the corresponding subgrade inclination degree can be predicted. Further, using the established classification mapping relationship, a plurality of corrected vehicle roll information collected in real time can be classified and processed, thereby obtaining a plurality of subgrade inclination analysis results reflecting the subgrade inclination conditions of different road sections or different positions of the same road section.

[0090] In order to further improve the detection accuracy, the second subgrade inclination detection result can be optimized and corrected by using a plurality of subgrade inclination analysis results obtained based on vehicle roll information. The specific optimization method can be weighted average method or fusion algorithm. Taking the weighted average method as an example, according to the reliability and accuracy of the vehicle roll information, different weights are given to the inclination analysis results from different sources, and then the weighted average value is calculated as the final correction result. After the above optimization and correction process, the third subgrade inclination detection result is obtained, which not only integrates the information from multiple data sources, but also effectively reduces the error caused by a single data source through intelligent processing of the machine learning model, thereby improving the accuracy and reliability of subgrade inclination detection.

[0091] Through the technical solutions of the above embodiments, the target detection optimization method based on multi-source data fusion provided by the present application solves the technical problems of insufficient accuracy, poor real-time performance, and information loss caused by single data in the subgrade inclination detection in the prior art. Through multi-source data fusion, the accuracy and real-time performance of subgrade inclination detection are improved, achieving high-precision and real-time subgrade inclination comprehensive detection and analysis, thereby ensuring the reliability of road maintenance and ensuring road safety.

[0092] Embodiment two, based on the same inventive concept as the target detection optimization method based on multi-source data fusion in the foregoing embodiments, as shown in Figure 2 The system provided by the present application comprises:

[0093] The subgrade data acquisition module 11 is configured to acquire, through the Internet of Things, a first subgrade inclination detection result of a target road region of a subgrade detected by an electronic inclinometer.

[0094] The roadbed inclination optimization correction module 12 is configured to collect roadbed point cloud data of the roadbed of the target road region obtained by laser scanning by the Internet of Things, perform roadbed fitting, obtain a roadbed inclination fitting result, and optimize and correct the first roadbed inclination detection result to obtain a second roadbed inclination detection result.

[0095] The vehicle image recognition module 13 is configured to collect a passing image sequence of a plurality of vehicles passing through the target road region by the Internet of Things, perform vehicle recognition, vehicle roll identification, and vehicle speed identification, and obtain a plurality of vehicle feature information, a plurality of vehicle roll information, and a plurality of vehicle speed information.

[0096] The vehicle roll optimization correction module 14 is configured to optimize and correct the plurality of vehicle roll information according to the plurality of vehicle speed information in combination with the plurality of vehicle feature information to obtain a plurality of corrected vehicle roll information.

[0097] The roadbed inclination detection result obtaining module 15 is configured to classify a plurality of roadbed inclination analysis results according to the plurality of corrected vehicle roll information, optimize and correct the second roadbed inclination detection result, and obtain a third roadbed inclination detection result.

[0098] Further, the roadbed data collection module 11 is further configured to perform the following steps:

[0099] An electronic inclinometer is configured at a roadbed detection position of the target road region, and is connected based on the Internet of Things, wherein the target road region is a curve region.

[0100] A plurality of roadbed inclination results of the roadbed obtained by the electronic inclinometer are collected by the Internet of Things, and a first roadbed inclination detection result is calculated and obtained, wherein each roadbed inclination result includes a roadbed inclination.

[0101] Further, the roadbed inclination optimization correction module 12 is further configured to perform the following steps:

[0102] Roadbed point cloud data of the roadbed of the target road obtained by laser scanning is collected by the Internet of Things.

[0103] A plurality of sample roadbed fitting anchor frames are obtained, wherein the roadbed inclinations in the plurality of sample roadbed fitting anchor frames are different.

[0104] The plurality of sample roadbed fitting anchor frames are used to perform traversal fitting in the roadbed point cloud data, and a sample roadbed fitting anchor frame with the largest roadbed fitting degree is obtained as the roadbed inclination fitting result.

[0105] Output the roadbed inclination of the sample roadbed fitting anchor frame in the roadbed inclination fitting result, and perform optimization correction calculation on the first roadbed inclination detection result to obtain a second roadbed inclination detection result.

[0106] Further, the roadbed inclination optimization correction module 12 is further configured to perform the following steps:

[0107] Randomly select a first sample roadbed fitting anchor frame from the plurality of sample roadbed fitting anchor frames.

[0108] Iterative fitting is performed in the roadbed point cloud data using the first sample roadbed fitting anchor frame. In each fitting, the proportion of the point cloud falling within the first sample roadbed fitting anchor frame in the roadbed point cloud data is calculated as the roadbed fitting degree. The fitting is finally converged to a number of times, and the maximum first roadbed fitting degree is output.

[0109] Continue to select sample roadbed fitting anchor frames from the plurality of sample roadbed fitting anchor frames for iterative fitting in the roadbed point cloud data, and obtain a plurality of roadbed fitting degrees.

[0110] Output the sample roadbed fitting anchor frame with the maximum roadbed fitting degree as the roadbed inclination fitting result.

[0111] Further, the vehicle image recognition module 13 is further configured to perform the following steps:

[0112] Vehicle recognition is performed on a plurality of passing images in a plurality of passing image sequences when a plurality of vehicles pass, and a plurality of identified vehicle sets are obtained. The vehicle recognizer is trained by collecting a sample passing image set and a sample identified vehicle set to perform vehicle recognition.

[0113] Based on the slowfast network, vehicle roll identification and vehicle speed identification are performed on the plurality of passing image sequences to obtain a plurality of vehicle roll information and a plurality of vehicle speed information.

[0114] Further, the vehicle image recognition module 13 is further configured to perform the following steps:

[0115] According to a first extraction step length, passing image extraction and down-sampling processing are respectively performed on the plurality of passing image sequences to obtain a plurality of first passing image sets.

[0116] According to a second extraction step length greater than the first extraction step length, passing image extraction is performed on the plurality of passing image sequences to obtain a plurality of second passing image sets.

[0117] Based on the historical monitoring data of the target road area, a plurality of sample first passing image sets, a plurality of sample second passing image sets, a sample vehicle roll information set, and a sample vehicle speed information set are collected.

[0118] The vehicle passing feature identifier is trained based on the slowfast network by using the plurality of sample first passing image sets, the plurality of sample second passing image sets, the sample vehicle roll information set and the sample vehicle speed information set.

[0119] The vehicle passing feature identifier is trained based on the slowfast network by using the plurality of sample first passing image sets, the plurality of sample second passing image sets, the sample vehicle roll information set and the sample vehicle speed information set.

[0120] Further, the vehicle roll optimization correction module 14 is further configured to perform the following steps:

[0121] Based on the vehicle roll test data of the same road area of the target road area, a sample vehicle feature information set, a sample vehicle speed information set and a sample speed roll information set are obtained.

[0122] The speed roll analyzer is trained by using the sample vehicle feature information set, the sample vehicle speed information set and the sample speed roll information set.

[0123] The speed roll analyzer is trained by using the sample vehicle feature information set, the sample vehicle speed information set and the sample speed roll information set.

[0124] The plurality of speed roll information is reduced in the plurality of vehicle roll information to obtain a plurality of corrected vehicle roll information.

[0125] Further, the roadbed tilt detection result obtaining module 15 is further configured to perform the following steps:

[0126] According to the monitoring data of the vehicle roll and the roadbed tilt, a sample corrected vehicle roll information set and a sample roadbed tilt analysis result set are collected, and each sample roadbed tilt analysis result includes a roadbed tilt degree.

[0127] A classification mapping relationship of the sample corrected vehicle roll information set and the sample roadbed tilt analysis result set is constructed.

[0128] According to the plurality of corrected vehicle roll information, a plurality of roadbed tilt analysis results are classified and obtained.

[0129] The second roadbed tilt detection result is optimized and corrected by using the plurality of roadbed tilt analysis results to obtain a third roadbed tilt detection result.

[0130] Through the foregoing detailed description of the target detection optimization method based on multi-source data fusion, those skilled in the art can clearly understand the target detection optimization system based on multi-source data fusion in the embodiment. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part is described in the method part.

[0131] The above description of disclosed embodiments enables one skilled in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A target detection optimization method based on multi-source data fusion, characterized in that, The method comprises: Collecting first subgrade inclination detection results of the subgrade of the target road region detected by the electronic inclinometer through the Internet of Things; Collecting subgrade point cloud data of the subgrade of the target road region scanned by the laser scanning device through the Internet of Things, performing subgrade fitting to obtain subgrade inclination fitting results, and optimizing and correcting the first subgrade inclination detection results to obtain second subgrade inclination detection results; Collecting passing image sequences of multiple vehicles passing through the target road region through the Internet of Things, performing vehicle identification, vehicle roll identification and vehicle speed identification to obtain multiple vehicle feature information, multiple vehicle roll information and multiple vehicle speed information; According to the multiple vehicle speed information and the multiple vehicle feature information, the multiple vehicle roll information is optimized and corrected to obtain multiple corrected vehicle roll information; According to the multiple corrected vehicle roll information, multiple subgrade inclination analysis results are classified to obtain third subgrade inclination detection results, and the second subgrade inclination detection results are optimized and corrected to obtain third subgrade inclination detection results; Wherein, the first subgrade inclination detection results of the subgrade of the target road region detected by the electronic inclinometer through the Internet of Things, comprising: In the subgrade detection position of the target road region, an electronic inclinometer is configured and connected based on the Internet of Things, wherein the target road region is a curve region; Collecting multiple subgrade inclination results of the subgrade obtained by the electronic inclinometer test through the Internet of Things, and calculating the first subgrade inclination detection results, wherein each subgrade inclination result comprises a subgrade inclination; Wherein, collecting the subgrade point cloud data of the subgrade of the target road region scanned by the laser scanning device through the Internet of Things, performing subgrade fitting to obtain subgrade inclination fitting results, and optimizing and correcting the first subgrade inclination detection results, comprising: Collecting the subgrade point cloud data of the subgrade of the target road scanned by the laser scanning device through the Internet of Things; Obtaining multiple sample subgrade fitting anchor frames, wherein the subgrade inclination in the multiple sample subgrade fitting anchor frames is different; Using the multiple sample subgrade fitting anchor frames, the subgrade point cloud data is iteratively fitted to obtain a sample subgrade fitting anchor frame with the maximum subgrade fitting degree as the subgrade inclination fitting result; Outputting the subgrade inclination of the sample subgrade fitting anchor frame in the subgrade inclination fitting result, and optimizing and correcting the first subgrade inclination detection result to obtain the second subgrade inclination detection result; Wherein, collecting passing image sequences of multiple vehicles passing through the target road region through the Internet of Things, performing vehicle identification, vehicle roll identification and vehicle speed identification to obtain multiple vehicle feature information, multiple vehicle roll information and multiple vehicle speed information, comprising: Vehicle identification is performed on multiple passing images in multiple passing image sequences when multiple vehicles pass through to obtain multiple identified vehicle sets, wherein a vehicle identifier is trained by collecting a sample passing image set and a sample identified vehicle set to perform vehicle identification; Based on the slowfast network, vehicle roll identification and vehicle speed identification are performed on the multiple passing image sequences to obtain multiple vehicle roll information and multiple vehicle speed information; Among them, based on the slowfast network, vehicle tilt recognition and vehicle speed recognition are performed on the multiple passed image sequences, including: According to the first extraction step size, the multiple passing image sequences are subjected to passing image extraction and downsampling processing respectively to obtain multiple first passing image sets; According to a second extraction step size greater than the first extraction step size, the multiple passed image sequences are extracted respectively to obtain multiple second passed image sets; Based on historical monitoring data of the target road area, multiple sample first pass image sets and multiple sample second pass image sets are collected, and sample vehicle side tilt information sets and sample vehicle speed information sets are obtained. Using the multiple sample first pass image sets, multiple sample second pass image sets, sample vehicle tilt information sets, and sample vehicle speed information sets, a vehicle pass feature recognizer is trained based on a slowfast network. The vehicle passing feature recognizer is used to input and recognize the multiple first passing image sets and multiple second passing image sets to obtain multiple vehicle tilt information and multiple vehicle speed information. 2.The target detection optimization method based on multi-source data fusion according to claim 1, characterized in that, Using the multiple sample roadbed fitting anchor frames, a traversal fitting is performed within the roadbed point cloud data to obtain the sample roadbed fitting anchor frame with the highest roadbed fitting degree, including: Randomly select the first sample roadbed fitting anchor frame from among the multiple sample roadbed fitting anchor frames; The first sample roadbed fitting anchor frame is used to iteratively fit within the roadbed point cloud data. In each fitting, the proportion of point cloud falling within the first sample roadbed fitting anchor frame in the roadbed point cloud data is calculated as the roadbed fitting degree. Finally, the number of fitting times is converged, and the maximum first roadbed fitting degree is output. Continue to iterate and select sample roadbed fitting anchor frames within the multiple sample roadbed fitting anchor frames in the roadbed point cloud data to obtain multiple roadbed fitting degrees. The sample roadbed fitting anchor frame with the highest roadbed fitting degree is output as the roadbed tilt fitting result. 3.The target detection optimization method based on multi-source data fusion according to claim 1, characterized in that, Based on the multiple vehicle speed information and the multiple vehicle feature information, the multiple vehicle roll information is optimized and corrected to obtain multiple corrected vehicle roll information, including: Based on vehicle roll test data from the same road area in the target road area, obtain a set of sample vehicle feature information, a set of sample vehicle speed information, and a set of sample speed roll information. The speed and roll analyzer is trained using the sample vehicle feature information set, sample vehicle speed information set, and sample speed roll information set. Using the speed roll analyzer, speed roll analysis is performed on the multiple vehicle speed information and the multiple vehicle feature information to obtain multiple speed roll information. By reducing the multiple speed roll information from the multiple vehicle roll information, multiple corrected vehicle roll information are obtained.

4. The target detection optimization method based on multi-source data fusion according to claim 1, characterized in that, Based on the multiple corrected vehicle roll information, multiple roadbed tilt analysis results are obtained by classification. The second roadbed tilt detection result is optimized and corrected to obtain the third roadbed tilt detection result, including: According to the monitoring data of vehicle roll and embankment inclination, a sample correction vehicle roll information set and a sample embankment inclination analysis result set are collected, each sample embankment inclination analysis result including embankment inclination; A classification mapping relationship of the sample correction vehicle roll information set and the sample embankment inclination analysis result set is constructed; According to the multiple correction vehicle roll information, multiple embankment inclination analysis results are obtained by classification; The multiple embankment inclination analysis results are used to perform optimization correction calculation on the second embankment inclination detection result, and a third embankment inclination detection result is obtained.

5. The target detection optimization system based on multi-source data fusion, characterized in that, The system for implementing the target detection optimization method based on multi-source data fusion of any one of claims 1-4 comprises: An embankment data acquisition module for acquiring first embankment inclination detection results of the embankment of the target road area detected by an electronic inclinometer through Internet of Things; An embankment inclination optimization correction module for acquiring embankment point cloud data of the embankment of the target road area scanned by a laser scanning device through Internet of Things, performing embankment fitting to obtain embankment inclination fitting results, and performing optimization correction on the first embankment inclination detection results to obtain second embankment inclination detection results; A vehicle image recognition module for acquiring a passing image sequence of multiple vehicles passing through the target road area through Internet of Things, performing vehicle recognition, vehicle roll recognition and vehicle speed recognition, and obtaining multiple vehicle characteristic information, multiple vehicle roll information and multiple vehicle speed information; A vehicle roll optimization correction module for performing optimization correction on the multiple vehicle roll information according to the multiple vehicle speed information combined with the multiple vehicle characteristic information to obtain multiple correction vehicle roll information; An embankment inclination detection result acquisition module for obtaining multiple embankment inclination analysis results by classification according to the multiple correction vehicle roll information, performing optimization correction on the second embankment inclination detection results, and obtaining a third embankment inclination detection result.

Images