Target detection system based on information entropy
Through the target detection system based on information entropy, grayscale and color information entropy are used to screen and merge target areas, combined with deep neural networks and anomaly detection, the detection difficulties of traditional methods in complex backgrounds are solved, and higher detection accuracy and abnormal target identification are achieved.
Patent Information
- Application Number
- CN202510826451.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-09-26
AI Technical Summary
Traditional target detection methods have difficulty effectively handling information confusion between target and non-target areas in complex backgrounds, and cannot accurately capture key features, resulting in missed detections or false detections.
An information entropy-based target detection system is adopted. Through data preprocessing, multimodal information fusion, information entropy calculation, target area screening and abnormal target detection, feature extraction and classification are performed in combination with deep neural networks. Grayscale and color information entropy are used to quantify regional uncertainty. Thresholds are set to screen target areas and merge them. Abnormal targets are judged in combination with an anomaly detection model.
Accurately distinguish between targets and backgrounds, enhance the ability to detect blurred and occluded targets, and improve detection accuracy and the ability to identify abnormal targets in complex scenarios.
Smart Images
Figure BDA0005458193230000031 
Figure BDA0005458193230000071 
Figure FDA0005458193220000021
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target detection, and in particular to a target detection system based on information entropy. Background Art
[0002] With the rapid development of artificial intelligence (AI), object detection, as one of the core tasks in computer vision, has found widespread application in numerous fields, including intelligent security, autonomous driving, and robotic navigation. Traditional object detection methods, such as those based on feature engineering and early deep learning approaches, can achieve good detection results in simple scenarios.
[0003] There are many challenges in actual target detection tasks. For example, in intelligent traffic monitoring scenarios, when strong sunlight directly hits the camera, the color, outline, and other features of the vehicle in the image will be blurred due to overexposure. Traditional detection algorithms may miss the vehicle or mistakenly identify the reflective area as the target. In forest fire warning systems, the textures of smoke and tree branches are similar, and the background interference is large. The detection algorithm has difficulty accurately distinguishing smoke targets, which can easily lead to false positives or omissions. In medical imaging diagnosis, the boundary between tumor tissue and normal tissue is blurred and the grayscale values are similar. Detection methods based on traditional feature extraction have difficulty accurately identifying tumor targets. The essence of these problems lies in the fact that traditional methods cannot effectively deal with the information confusion between target and non-target areas in complex backgrounds, and cannot accurately capture the key features of the target.
[0004] Based on the above, a target detection system based on information entropy is invented. Summary of the Invention
[0005] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:
[0006] An object detection system based on information entropy includes the following specific steps:
[0007] S1, data preprocessing: first collect and annotate image data related to the target detection task, then perform image enhancement and normalization;
[0008] S2, multimodal information fusion: aligning and fusing features of data from different modalities;
[0009] S3, information entropy calculation: first divide the image into blocks, then for each sub-block, calculate its grayscale information entropy. If the image is in color, the color information entropy is also calculated;
[0010] S4, target area screening: First, an information entropy threshold is set based on the overall information entropy distribution of the image to screen out areas that may contain targets. Then, the sub-blocks that have been initially screened are merged to form a more complete target candidate area.
[0011] S5, feature extraction and classification: First, extract the deep features of the target candidate area, then fuse the calculated information entropy with the extracted features, and then classify the target candidate area to complete the target detection task;
[0012] S6, abnormal target detection: Determine whether the information entropy feature is an abnormal target based on the anomaly detection model.
[0013] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S1 are as follows:
[0014] S11, Image Acquisition and Annotation: First, collect image data related to the target detection task, covering various complex scenes. Then use professional annotation tools to accurately annotate the targets in the image. The annotation information includes the target category and location coordinates.
[0015] S12, image enhancement: first enhance the contrast of the image; then remove the noise interference in the image to improve the image quality;
[0016] S13, normalization processing: adjust the size of the image to a fixed size and normalize the pixel values of the image to facilitate subsequent model training and calculation.
[0017] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S2 are as follows:
[0018] S21, Multimodal Data Acquisition and Preparation: In the case of target detection based on information entropy, data of different modalities are acquired according to specific application requirements. After data acquisition, preliminary preprocessing of each modality is performed to ensure data quality and availability.
[0019] S22, multimodal data alignment: performing a data alignment operation on the multimodal data so that the different modal data can accurately correspond to the same state of the target, the data alignment operation includes time alignment and spatial alignment;
[0020] S23, multimodal data fusion: performing a fusion operation on the aligned multimodal data according to different fusion strategies; the fusion operation includes early fusion, late fusion, and hybrid fusion;
[0021] S24, post-fusion data processing: further processing the fused data to meet the input requirements of subsequent information entropy calculation and target detection model, and the further processing includes data normalization, data enhancement and format conversion.
[0022] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S3 are as follows:
[0023] S31, image segmentation: dividing the fused image into sub-blocks of equal size, where the size of the sub-blocks can be adjusted according to the image resolution and computing resources;
[0024] Grayscale information entropy calculation: For each sub-block, calculate its grayscale information entropy. The grayscale information entropy calculation formula is:
[0025]
[0026] Where L is the number of gray levels, p(i) represents the probability of a pixel with gray value i appearing in a sub-block, and grayscale information entropy reflects the uncertainty of grayscale distribution within a sub-block. The higher the information entropy, the more complex the grayscale changes within the sub-block, and the more likely it is to contain target information.
[0027] S32, color information entropy calculation: If the image is a color image, the image will first be converted from the RGB color space to the HSV color space, and the information entropy of the three channels of hue H, saturation S, and brightness V will be calculated respectively. Then, the information entropy of the three channels will be weightedly fused to obtain the color information entropy. The color information entropy can further describe the color distribution characteristics of the image sub-blocks and assist in distinguishing the target from the background.
[0028] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S4 are as follows:
[0029] S41, threshold setting: First, set an information entropy threshold T based on the overall information entropy distribution of the image. Then, compare the information entropy of each sub-block with the information entropy threshold T. Sub-blocks with information entropy greater than the threshold are preliminarily determined to be areas that may contain targets. The calculation formula of the information entropy threshold T is:
[0030]
[0031] in, is the mean of the information entropy of all sub-blocks, k is a coefficient ranging from 1 to 3, and σ is the standard deviation of the information entropy of all sub-blocks;
[0032] S42, region merging: The connectivity-based region merging algorithm performs region merging on the preliminarily screened sub-blocks to merge adjacent sub-blocks with high information entropy into a large region. This allows sub-blocks with similar features to be connected to form a more complete target candidate region.
[0033] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S5 are as follows:
[0034] S51, deep feature extraction: input the screened target candidate area into the deep neural network to extract the deep features of the target;
[0035] S52, Information Entropy Fusion: The calculated information entropy is used as an additional feature and fused with the features extracted by the deep neural network to enable the model to better utilize the information contained in the information entropy to distinguish between the target and the background, thereby improving the accuracy of detection;
[0036] S53, target classification and positioning: Using the fused features, the classifier is used to classify the target candidate area to determine whether it is a real target and the category it belongs to. At the same time, the regressor is used to predict the position coordinates and size information of the target to achieve accurate positioning of the target. At this point, the target detection task is completed.
[0037] As a preferred solution of the target detection system based on information entropy described in the present invention, the specific steps of S6 are as follows:
[0038] S61, Data Feature Analysis and Normal Pattern Modeling: After completing target classification and positioning, first collect labeled normal target data, then extract the information entropy features of the normal target data, and then use statistical methods to analyze the feature distribution patterns of the normal target data and build a feature distribution model for normal targets;
[0039] S62, Anomaly Detection Model Construction and Training: First, based on the constructed normal target feature distribution model, select the corresponding anomaly detection algorithm to build the model. Then, use the collected normal target data and known abnormal target data to train and optimize the model, and adjust the model parameters so that it can accurately distinguish normal targets from abnormal targets.
[0040] S63, abnormal target determination: The target data after target classification and positioning is input into the trained anomaly detection model, so that the model calculates the probability or score of the target being abnormal based on the pre-set judgment rules and the information entropy characteristics and other characteristics of the target, and marks each target based on the calculation result to determine whether it is an abnormal target;
[0041] S64, abnormal target feedback and recording: For targets determined to be abnormal, relevant information is promptly fed back to the system administrator. The feedback information includes images or video clips of the abnormal target, the target's location coordinates, category information, and anomaly score, so that relevant personnel or modules can quickly understand the abnormal situation. At the same time, all relevant data of the abnormal target, including original data, extracted feature data, and judgment results, are recorded and stored in the database for subsequent model optimization, helping the model learn more features of abnormal targets and improve the ability to detect abnormal targets.
[0042] Compared with existing technologies:
[0043] 1. Accurately distinguish between target and background: Traditional methods have difficulty distinguishing between target and background under complex background interference. For example, in forest fire warning systems, the similar textures of smoke and tree branches and leaves often cause detection algorithms to misjudge. The information entropy-based method can quantify the information uncertainty of each area by calculating the grayscale information entropy and color information entropy (color image) of the image area. Areas with high information entropy have complex grayscale or color changes and are more likely to contain targets. By setting appropriate thresholds to filter high information entropy areas and merging areas, the target area and background area can be effectively separated, greatly reducing the impact of background interference on the detection results. For example, in intelligent traffic monitoring, even if the vehicle's features are blurred due to overexposure, this method can still identify the area where the target vehicle is located based on the information entropy characteristics.
[0044] 2. Enhance the ability to detect blurred and occluded targets: In actual scenarios, targets are often blurred or occluded, and traditional detection methods find it difficult to accurately capture features. For example, when the boundary between tumor tissue and normal tissue is blurred in medical imaging diagnosis, traditional methods have difficulty in identification. The information entropy-based method uses information entropy to measure the uncertainty of target features, and when extracting features in deep neural networks, information entropy is fused as an additional feature. This enables the model to adaptively adjust the degree of attention to different features, focusing on key information, thereby improving the ability to detect blurred and occluded targets. For example, in an autonomous driving scenario, when a vehicle is partially obscured by a billboard, this method can still accurately detect the vehicle target with the assistance of information entropy and combined with deep features. DETAILED DESCRIPTION
[0045] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below.
[0046] The present invention provides a target detection system based on information entropy, which includes the following specific steps:
[0047] S1, data preprocessing: first collect and annotate image data related to the target detection task, then perform image enhancement and normalization;
[0048] The specific steps of S1 are as follows:
[0049] S11, Image Acquisition and Annotation: First, collect image data related to the target detection task, covering various complex scenes. Then use professional annotation tools to accurately annotate the targets in the image. The annotation information includes the target category and location coordinates.
[0050] S12, image enhancement: first enhance the contrast of the image; then remove the noise interference in the image to improve the image quality;
[0051] S13, normalization processing: resize the image to a fixed size and normalize the pixel values of the image to facilitate subsequent model training and calculation;
[0052] S2, multimodal information fusion: aligning and fusing features of data from different modalities;
[0053] The specific steps of S2 are as follows:
[0054] S21, Multimodal Data Acquisition and Preparation: In the case of target detection based on information entropy, data of different modalities are acquired according to specific application requirements. After data acquisition, preliminary preprocessing of each modality is performed to ensure data quality and availability.
[0055] S22, multimodal data alignment: performing a data alignment operation on the multimodal data so that the different modal data can accurately correspond to the same state of the target, the data alignment operation includes time alignment and spatial alignment;
[0056] S23, multimodal data fusion: performing a fusion operation on the aligned multimodal data according to different fusion strategies; the fusion operation includes early fusion, late fusion, and hybrid fusion;
[0057] The early fusion step involves directly merging data from different modalities. For image and point cloud data, the point cloud data can be projected and transformed and then joined with the image data as a new channel. For example, the point cloud data can be projected onto an RGB image to generate a depth map containing the point cloud information. The depth map and RGB image are then merged into four-channel data, which serves as the data for subsequent information entropy calculation and model input. This approach fully utilizes the complementarity of the underlying data, allowing the model to learn the comprehensive information of multimodal data during the feature extraction stage.
[0058] The late fusion step is to perform independent feature extraction and target detection on data from different modalities, and then fuse the detection results of different modalities. For example, first use a convolutional neural network to detect targets on RGB images to obtain detection results in the image modality. Then use a point cloud processing network to detect the lidar point cloud data to obtain detection results in the point cloud modality. Finally, the detection results of the two modalities are fused through strategies such as voting and weighted averaging. For example, different weights are assigned to the detection results of the two modalities based on their performance in different scenarios, and the weighted results are combined to obtain the final target detection result.
[0059] Hybrid fusion: Combining early and late fusion methods, fusion operations are performed at different stages of data processing. For example, some modal data is first early fused to extract comprehensive features; other modal data are then processed independently and the results of different parts are then fused in the subsequent feature extraction or decision-making stages. This approach can flexibly combine the advantages of different fusion strategies and adapt to complex and changing application scenarios.
[0060] S24, post-fusion data processing: further processing the fused data to meet the input requirements of subsequent information entropy calculation and target detection model, wherein the further processing includes data normalization, data enhancement and format conversion;
[0061] By setting up multimodal information fusion, it is possible to fully utilize the complementary advantages of different modal data to provide richer and more comprehensive target information. For example, in nighttime autonomous driving scenarios, image data may not be able to clearly present the target due to insufficient light, while lidar point cloud data can accurately obtain the target's three-dimensional structural information. After the two are fused, information entropy calculation can be used to more accurately identify and locate the target, effectively improving target detection capabilities in complex scenarios and compensating for the limitations of single modal data.
[0062] S3, information entropy calculation: first divide the image into blocks, then for each sub-block, calculate its grayscale information entropy. If the image is in color, the color information entropy is also calculated;
[0063] The specific steps of S3 are as follows:
[0064] S31, image segmentation: dividing the fused image into sub-blocks of equal size, where the size of the sub-blocks can be adjusted according to the image resolution and computing resources;
[0065] Grayscale information entropy calculation: For each sub-block, calculate its grayscale information entropy. The grayscale information entropy calculation formula is:
[0066]
[0067] Where L is the number of gray levels, p(i) represents the probability of a pixel with gray value i appearing in a sub-block, and grayscale information entropy reflects the uncertainty of grayscale distribution within a sub-block. The higher the information entropy, the more complex the grayscale changes within the sub-block, and the more likely it is to contain target information.
[0068] S32, color entropy calculation: If the image is a color image, the image is first converted from the RGB color space to the HSV color space, and the information entropy of the three channels of hue H, saturation S, and brightness V are calculated respectively. The information entropy of the three channels is then weighted and fused to obtain the color information entropy. The color information entropy can further describe the color distribution characteristics of the image sub-blocks and assist in distinguishing the target from the background.
[0069] S4, target area screening: First, an information entropy threshold is set based on the overall information entropy distribution of the image to screen out areas that may contain targets. Then, the sub-blocks that have been initially screened are merged to form a more complete target candidate area.
[0070] The specific steps of S4 are as follows:
[0071] S41, threshold setting: First, set an information entropy threshold T based on the overall information entropy distribution of the image. Then, compare the information entropy of each sub-block with the information entropy threshold T. Sub-blocks with information entropy greater than the threshold are preliminarily determined to be areas that may contain targets. The calculation formula of the information entropy threshold T is:
[0072]
[0073] in, is the mean of the information entropy of all sub-blocks, k is a coefficient ranging from 1 to 3, and σ is the standard deviation of the information entropy of all sub-blocks;
[0074] S42, region merging: A connectivity-based region merging algorithm performs region merging on the initially screened sub-blocks to merge adjacent sub-blocks with high information entropy into a large region. This allows sub-blocks with similar features to be connected to form a more complete target candidate region.
[0075] S5, feature extraction and classification: First, extract the deep features of the target candidate area, then fuse the calculated information entropy with the extracted features, and then classify the target candidate area to complete the target detection task;
[0076] The specific steps of S5 are as follows:
[0077] S51, deep feature extraction: input the screened target candidate area into the deep neural network to extract the deep features of the target;
[0078] S52, Information Entropy Fusion: The calculated information entropy is used as an additional feature and fused with the features extracted by the deep neural network to enable the model to better utilize the information contained in the information entropy to distinguish between the target and the background, thereby improving the accuracy of detection;
[0079] S53, Target Classification and Localization: Using the fused features, the classifier classifies the target candidate area to determine whether it is a real target and the category it belongs to. At the same time, the regressor is used to predict the target's position coordinates and size information to achieve accurate positioning of the target. At this point, the target detection task is completed;
[0080] S6, abnormal target detection: determine whether the information entropy feature is an abnormal target based on the anomaly detection model;
[0081] The specific steps of S6 are as follows:
[0082] S61, Data Feature Analysis and Normal Pattern Modeling: After completing target classification and positioning, first collect labeled normal target data, then extract the information entropy features of the normal target data, and then use statistical methods to analyze the feature distribution patterns of the normal target data and build a feature distribution model for normal targets;
[0083] S62, anomaly detection model construction and training: First, based on the constructed normal target feature distribution model, select the corresponding anomaly detection algorithm to build the model. A distance-based algorithm (such as the local outlier factor LOF algorithm) can be used to determine whether the target data point is an outlier by calculating the distance between the target data point and the normal target data point. A density-based algorithm can also be used to analyze the density around the target data point. If the density is significantly lower than the normal target area, it is determined to be an anomaly. A deep learning model, such as an auto-encoder, can also be used to train the model to learn the feature representation of normal targets. When abnormal target data is input, the model reconstruction error will increase significantly, thereby identifying anomalies. Then, the model is trained and optimized using the collected normal target data and known abnormal target data, and the model parameters are adjusted so that it can accurately distinguish normal targets from abnormal targets.
[0084] S63, abnormal target determination: The target data after target classification and positioning is input into the trained anomaly detection model, so that the model calculates the probability or score of the target being abnormal based on the pre-set judgment rules and the information entropy characteristics and other characteristics of the target, and marks each target based on the calculation result to determine whether it is an abnormal target;
[0085] S64, Abnormal Target Feedback and Recording: For targets determined to be abnormal, relevant information is promptly fed back to system administrators. The feedback information includes images or video clips of the abnormal target, the target's location coordinates, category information, and anomaly score, so that relevant personnel or modules can quickly understand the abnormal situation. At the same time, all relevant data of the abnormal target, including raw data, extracted feature data, and judgment results, is recorded and stored in a database for subsequent model optimization, helping the model learn more features of abnormal targets and improve its detection capabilities.
[0086] By setting up abnormal target detection, we can promptly detect unusual objects in scenarios, such as identifying people carrying unusual items during airport security checks or detecting products with unusual shapes or features during industrial production inspections. Prompt feedback and recording not only helps ensure safety and improve production quality, but also provides special samples for model optimization, enhancing the model's ability to detect rare or unusual objects.
[0087] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A target detection system based on information entropy, characterized in that: The specific steps are as follows: S1, data preprocessing: first collect and annotate image data related to the target detection task, then perform image enhancement and normalization; S2, multimodal information fusion: aligning and fusing features of data from different modalities; S3, information entropy calculation: first divide the image into blocks, then for each sub-block, calculate its grayscale information entropy. If the image is in color, the color information entropy is also calculated; S4, target area screening: First, an information entropy threshold is set based on the overall information entropy distribution of the image to screen out areas that may contain targets. Then, the sub-blocks that have been initially screened are merged to form a more complete target candidate area. S5, feature extraction and classification: First, extract the deep features of the target candidate area, then fuse the calculated information entropy with the extracted features, and then classify the target candidate area to complete the target detection task; S6, abnormal target detection: Determine whether the information entropy feature is an abnormal target based on the anomaly detection model.
2. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S1 are as follows: S11, Image Acquisition and Annotation: First, collect image data related to the target detection task, covering various complex scenes. Then use professional annotation tools to accurately annotate the targets in the image. The annotation information includes the target category and location coordinates. S12, image enhancement: first enhance the contrast of the image; then remove the noise interference in the image to improve the image quality; S13, normalization processing: adjust the size of the image to a fixed size and normalize the pixel values of the image to facilitate subsequent model training and calculation.
3. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S2 are as follows: S21, Multimodal Data Acquisition and Preparation: In the case of target detection based on information entropy, data of different modalities are acquired according to specific application requirements. After data acquisition, preliminary preprocessing of each modality is performed to ensure data quality and availability. S22, multimodal data alignment: performing a data alignment operation on the multimodal data so that the different modal data can accurately correspond to the same state of the target, the data alignment operation includes time alignment and spatial alignment; S23, multimodal data fusion: performing a fusion operation on the aligned multimodal data according to different fusion strategies; the fusion operation includes early fusion, late fusion, and hybrid fusion; S24, post-fusion data processing: further processing the fused data to meet the input requirements of subsequent information entropy calculation and target detection model, and the further processing includes data normalization, data enhancement and format conversion.
4. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S3 are as follows: S31, image segmentation: dividing the fused image into sub-blocks of equal size, where the size of the sub-blocks can be adjusted according to the image resolution and computing resources; Grayscale information entropy calculation: For each sub-block, calculate its grayscale information entropy. The grayscale information entropy calculation formula is: Where L is the number of gray levels, p(i) represents the probability of a pixel with gray value i appearing in a sub-block, and grayscale information entropy reflects the uncertainty of grayscale distribution within a sub-block. The higher the information entropy, the more complex the grayscale changes within the sub-block, and the more likely it is to contain target information. S32, color information entropy calculation: If the image is a color image, the image will first be converted from the RGB color space to the HSV color space, and the information entropy of the three channels of hue H, saturation S, and brightness V will be calculated respectively. Then, the information entropy of the three channels will be weightedly fused to obtain the color information entropy. The color information entropy can further describe the color distribution characteristics of the image sub-blocks and assist in distinguishing the target from the background.
5. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S4 are as follows: S41, threshold setting: First, set an information entropy threshold T based on the overall information entropy distribution of the image. Then, compare the information entropy of each sub-block with the information entropy threshold T. Sub-blocks with information entropy greater than the threshold are preliminarily determined to be areas that may contain targets. The calculation formula of the information entropy threshold T is: in, is the mean of the information entropy of all sub-blocks, k is a coefficient ranging from 1 to 3, and σ is the standard deviation of the information entropy of all sub-blocks; S42, region merging: The connectivity-based region merging algorithm performs region merging on the preliminarily screened sub-blocks to merge adjacent sub-blocks with high information entropy into a large region. This allows sub-blocks with similar features to be connected to form a more complete target candidate region.
6. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S5 are as follows: S51, deep feature extraction: input the screened target candidate area into the deep neural network to extract the deep features of the target; S52, Information Entropy Fusion: The calculated information entropy is used as an additional feature and fused with the features extracted by the deep neural network to enable the model to better utilize the information contained in the information entropy to distinguish between the target and the background, thereby improving the accuracy of detection; S53, target classification and positioning: Using the fused features, the classifier is used to classify the target candidate area to determine whether it is a real target and the category it belongs to. At the same time, the regressor is used to predict the position coordinates and size information of the target to achieve accurate positioning of the target. At this point, the target detection task is completed.
7. The target detection system based on information entropy according to claim 1, characterized in that: The specific steps of S6 are as follows: S61, data feature analysis and normal pattern modeling: After completing target classification and positioning, first collect the labeled normal target data, and then extract its information entropy features for the normal target data. Then, statistical methods are used to analyze the characteristic distribution patterns of normal target data and construct a characteristic distribution model of normal targets; S62, Anomaly Detection Model Construction and Training: First, based on the constructed normal target feature distribution model, select the corresponding anomaly detection algorithm to build the model. Then, use the collected normal target data and known abnormal target data to train and optimize the model, and adjust the model parameters so that it can accurately distinguish normal targets from abnormal targets. S63, abnormal target determination: The target data after target classification and positioning is input into the trained anomaly detection model, so that the model calculates the probability or score of the target being abnormal based on the pre-set judgment rules and the information entropy characteristics and other characteristics of the target, and marks each target based on the calculation result to determine whether it is an abnormal target; S64, abnormal target feedback and recording: For targets determined to be abnormal, relevant information is promptly fed back to the system administrator. The feedback information includes images or video clips of the abnormal target, the target's location coordinates, category information, and anomaly score, so that relevant personnel or modules can quickly understand the abnormal situation. At the same time, all relevant data of the abnormal target, including original data, extracted feature data, and judgment results, are recorded and stored in the database for subsequent model optimization, helping the model learn more features of abnormal targets and improve the ability to detect abnormal targets.