Police service management system and method based on image recognition technology

Abstract

The invention relates to the technical field of police service management, discloses a police service management system and method based on an image recognition technology, and aims to improve the police service management efficiency and accuracy. The system comprises a data receiving module, a data preprocessing module, a target detection module, a behavior analysis module and an information integration module. A monitoring video stream is received, data is preprocessed to form an image data set, and then a specific target is detected by using a deep convolutional neural network and a track sequence is output. And carrying out behavior analysis on the target track by adopting a long-short-term memory network, and identifying an abnormal mode. And finally, integrating the detection result and the behavior analysis data, and generating a police service management report. According to the invention, intelligent processing of monitoring videos is realized, police service departments are assisted to respond to abnormal conditions in time, and urban public security is maintained. Through the deep learning technology, the accuracy of target detection and behavior recognition is improved, and powerful support is provided for modern police service management.

Description

Technical Field

[0001] The present invention relates to the technical field of police management, and in particular to a police management system and method based on image recognition technology. Background Art

[0002] With the rapid development of society and the continuous advancement of science and technology, image recognition technology has become an indispensable part of modern information construction, especially in the field of police management, showing great application potential. In recent years, the rapid development of deep learning and artificial intelligence technology has provided strong technical support for the innovation of image recognition technology and brought new solutions and ideas to police management. Traditional police monitoring systems mainly rely on manual monitoring and video playback, which is not only inefficient, but also easily affected by human factors, and it is difficult to meet the requirements of modern police management for efficiency and accuracy. With the surge in the number of surveillance cameras, the massive amount of video data has brought a heavy burden to police officers, making the extraction and analysis of key information extremely difficult.

[0003] In order to overcome these challenges, intelligent monitoring systems based on image recognition have emerged. Such systems can automatically analyze surveillance videos using deep learning algorithms to achieve accurate detection, tracking, and behavior analysis of specific targets. Through advanced algorithms such as deep convolutional neural networks (CNNs), the system can accurately identify target objects in images. However, although existing intelligent monitoring systems have achieved remarkable results in target detection and tracking, they are still insufficient in the recognition and analysis of complex behavior patterns. Traditional behavior analysis methods often rely on manually designed feature extraction and rule judgment, which is difficult to adapt to complex and changing monitoring scenarios. At the same time, with the continuous renovation of criminal methods, police management has put forward higher requirements for the accuracy and real-time nature of behavior analysis. Summary of the invention

[0004] The purpose of the present invention is to provide a police management system and method based on image recognition technology to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a police management system based on image recognition technology, the system comprising:

[0006] A data receiving module, used for receiving video stream and image data from a surveillance camera;

[0007] The data preprocessing module performs format conversion, denoising, enhancement and feature extraction on the received data to form a preprocessed image data set;

[0008] The target detection module builds a target detection model based on a deep convolutional neural network algorithm. The model is trained using a preprocessed image dataset to detect specific targets in the image and output a target trajectory sequence, each of which contains the target's continuous position information and timestamp;

[0009] The behavior analysis module builds a behavior analysis model based on the long short-term memory network algorithm. The model uses the target trajectory sequence output by the target detection model to analyze and identify the abnormal behavior pattern of the target;

[0010] The information integration module integrates the output results of the target detection model and the behavior analysis model to generate a police management report.

[0011] Preferably, the target detection model adopts a YOLO network structure, and its network structure includes:

[0012] Input layer: receives preprocessed image data, and the image size is unified to [H, W, C], where H is the height, W is the width, and C is the number of channels;

[0013] Convolutional layer: multi-layer convolution operation is adopted, each convolution layer is followed by a Leaky ReLU activation function, the convolution kernel size is set to a combination of 3x3 and 1x1, and the step size is set to 1;

[0014] Feature fusion layer: fuses feature maps of different scales through the feature pyramid network structure;

[0015] Prediction layer: The fused feature map is divided into grids, each grid predicts a fixed number of bounding boxes and category probabilities, and non-maximum suppression is used to filter the final detection results;

[0016] Output layer: Outputs the category, location, and confidence of the detected target, as well as the target trajectory sequence generated based on continuous frame data, which contains the continuous location information and timestamp of the target.

[0017] Preferably, the training step of the target detection model includes:

[0018] Initialize the parameters of the YOLO network model, including convolution kernel weights, bias terms, and learning rates;

[0019] Divide the preprocessed image dataset into training set, validation set and test set;

[0020] The YOLO network model is iteratively trained using the training set. In each iteration, the forward propagation calculates the model output, and the loss between the predicted value and the true value is calculated according to the loss function. The loss function includes coordinate loss, confidence loss, and classification loss. The calculation formula is:

[0021] Loss = λ_coord Loss_coord + λ_noobj Loss_noobj + λ_obj Loss_obj + Loss_class

[0022] Among them, Loss_coord is the coordinate loss, Loss_noobj is the confidence loss of the non-target box, Loss_obj is the confidence loss of the target box, and Loss_class is the classification loss; λ_coord, λ_noobj, λ_obj and Loss_class are the weights of Loss_coord, Loss_noobj, Loss_obj and Loss_class respectively;

[0023] According to the loss value, the model parameters are updated through the back-propagation algorithm and the Adam optimizer;

[0024] Repeat the training steps until the loss converges or the preset number of iterations is reached;

[0025] The trained YOLO network model is verified using the validation set to evaluate the detection accuracy and recall rate of the model. When the model performance reaches the preset standard, the test set is used for the final test, and the model parameters are saved to obtain the trained target detection model.

[0026] Preferably, the network structure of the behavior analysis model includes:

[0027] Input layer: receives the target trajectory sequence output by the target detection model. Each sequence contains the continuous position information and timestamp of the target.

[0028] LSTM layer: contains multiple layers of LSTM units. Each layer of LSTM units processes the input sequence and captures the dependencies in the time dimension.

[0029] Fully connected layer: The output features of the LSTM layer are classified through the fully connected layer to identify the abnormal behavior patterns of the target;

[0030] Output layer: Outputs the abnormal behavior category and confidence of the target.

[0031] Preferably, the step of training the behavior analysis model includes:

[0032] Use the target trajectory sequence output by the target detection model as the training set and select the abnormal behavior category as the label;

[0033] Sequence the training set and divide the continuous target trajectory into sequence segments of fixed length;

[0034] The LSTM network model is iteratively trained using sequence segments. In each iteration, the model output is calculated by forward propagation, and the loss between the predicted value and the true value is calculated according to the cross entropy loss function.

[0035] According to the loss value, the model parameters are updated through the back-propagation algorithm and the RMSprop optimizer;

[0036] Repeat the training steps until the loss converges or the preset number of iterations is reached;

[0037] The trained LSTM network model is verified using the validation set to evaluate the recognition accuracy and generalization ability of the model. When the model performance reaches the preset standard, the model parameters are saved to obtain the trained behavior analysis model.

[0038] Preferably, the information integration module uses a rule engine to integrate the output results of the target detection model and the behavior analysis model, and the specific steps include:

[0039] The output of the target detection model is defined as a detection event, which includes the target category, location, time and confidence.

[0040] The output of the behavior analysis model is defined as a behavior event, which includes the behavior category, target ID, time period, and confidence level.

[0041] According to the preset rules, the detection events and behavior events are correlated and analyzed;

[0042] Generate police management reports containing relevant incidents, risk levels and recommended actions.

[0043] Preferably, the step of performing correlation analysis on the detection event and the behavior event according to a preset rule further comprises:

[0044] Setting an association condition, wherein the association condition is based on at least one of the target category, location, time, and confidence in the detection event, and the behavior category, target ID, time period, and confidence in the behavior event;

[0045] Matching detection events and behavior events that meet the association condition to form an associated event;

[0046] Conduct a risk assessment on the associated events and determine the risk level based on the risk assessment results.

[0047] Preferably, the step of generating a police management report including associated events, risk levels and recommended measures further comprises:

[0048] According to the associated events and risk levels, select corresponding recommended measures from the preset recommended measures library;

[0049] Combining the associated events, risk levels and selected recommended actions into a police management report;

[0050] The police management report is output for reference by police management personnel in making decisions.

[0051] Preferably, the data preprocessing module uses a principal component analysis (PCA) algorithm to extract features, and the specific steps include:

[0052] Calculate the covariance matrix of image data;

[0053] Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues;

[0054] Select the eigenvectors corresponding to the first k largest eigenvalues ​​to form the feature space;

[0055] The original image data is projected into the feature space to obtain the reduced-dimensional feature data as the model input.

[0056] Preferably, a police management method based on image recognition technology comprises:

[0057] Step 1: Receive video stream and image data from the surveillance camera;

[0058] Step 2: Perform format conversion, denoising, enhancement and feature extraction on the received data to form a preprocessed image data set;

[0059] Step 3: Build a target detection model based on a deep convolutional neural network algorithm and use the preprocessed image dataset for training to detect specific targets in the image and output a trajectory sequence of the target. Each sequence contains the continuous position information and timestamp of the target.

[0060] Step 4: Build a behavior analysis model based on the long short-term memory network algorithm, and use the target trajectory sequence output by the target detection model to analyze and identify the abnormal behavior pattern of the target;

[0061] Step 5: Integrate the output results of the target detection model and the behavior analysis model to generate a police management report.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] Traditional police surveillance systems are deficient in the recognition and analysis of complex behavior patterns, often relying on manually designed feature extraction and rule judgment. This system integrates advanced behavior analysis modules to automatically recognize and analyze complex behavior patterns in surveillance videos, adapt to complex and changing surveillance scenarios, and meet the high requirements of police management for behavior analysis.

[0064] With the surge in the number of surveillance cameras, the massive amount of video data has brought a heavy burden to police officers. This system can automatically analyze surveillance videos, extract key information, and generate police management reports, which greatly reduces the workload of police officers and improves work efficiency. The system can monitor the behavior of the target in real time. Once an abnormal behavior pattern is found, it can immediately trigger the early warning mechanism, allowing police officers to respond quickly and effectively prevent and combat criminal behavior. The information integration module integrates the results of target detection and behavior analysis into a police management report, providing comprehensive and accurate data support for police decision-making, helping police departments make more scientific and reasonable decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a working principle diagram of the police management system based on image recognition technology according to the present invention;

[0066] Figure 2 This is a diagram of the training steps for the target detection model;

[0067] Figure 3 A flowchart for generating a police management report to integrate the output of the target detection model and the behavior analysis model. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] See also Figure 1-3 The present invention provides a technical solution: a police management system based on image recognition technology, the system comprising:

[0070] Data receiving module: The data receiving module is responsible for receiving video streams and image data from various surveillance cameras. This module establishes a connection with the surveillance camera through standard network communication protocols such as RTMP, HLS or HTTP Live Streaming to ensure stable and efficient reception of real-time video stream data. At the same time, this module also supports reading static image data from storage devices to meet the needs of retrospective analysis of historical data.

[0071] Data preprocessing module: The data preprocessing module performs a series of preprocessing operations on the received video stream and image data. First, format conversion is performed to uniformly convert data from different sources into a format suitable for subsequent processing, such as JPEG or PNG. Next, denoising is performed to remove noise and interference in the image through filtering algorithms to improve image quality. Then, image enhancement is performed to make the image clearer and easier to analyze by adjusting parameters such as brightness, contrast, and color balance. Finally, feature extraction is performed to extract key features in the image using algorithms such as edge detection and corner detection to provide basic data for subsequent target detection and behavior analysis.

[0072] Target detection module: The target detection module uses a deep convolutional neural network (CNN) algorithm to build a target detection model. The model is trained using a preprocessed image dataset and automatically learns feature representations in images through a combination of convolutional layers, pooling layers, and fully connected layers. During the training process, the model parameters are continuously optimized through the back-propagation algorithm to improve the accuracy of target detection. After training, the model can automatically detect specific targets in the image, such as pedestrians, vehicles, etc., and output the target's trajectory sequence. Each trajectory sequence contains the target's continuous position information and corresponding timestamps, providing data support for subsequent behavior analysis.

[0073] Behavior Analysis Module: The behavior analysis module builds a behavior analysis model based on the long short-term memory network (LSTM) algorithm. The model receives the target trajectory sequence output by the target detection module as input and captures the temporal dependencies in the sequence through the memory unit of the LSTM network. During the training process, the model parameters are optimized by minimizing the prediction error so that it can accurately identify the abnormal behavior patterns of the target. Once the training is completed, the model can automatically analyze and identify abnormal behaviors in the target trajectory sequence, such as sudden acceleration, sharp turns, etc., and provide timely abnormal alerts for police officers.

[0074] Information integration module: The information integration module integrates the output results of the target detection model and the behavior analysis model. The trajectory sequence output by the target detection model is visualized to generate intuitive visualization results such as trajectory maps or heat maps. Combined with the abnormal behavior patterns identified by the behavior analysis model, a detailed police management report is generated. The report contains the target's trajectory information, abnormal behavior descriptions, and possible security risks, providing comprehensive and accurate data support for police officers to help them make decisions quickly and take corresponding actions.

[0075] The present invention will be further described below in conjunction with Examples 1 to 3:

[0076] Embodiment 1:

[0077] The target detection model adopts the YOLO (You Only Look Once) network structure, which includes:

[0078] Input layer: Before the model starts processing, the raw image data needs to go through a preprocessing step to ensure the uniformity of the image size. Preprocessing includes but is not limited to operations such as resizing and normalizing the image. After preprocessing, the image data has a uniform size of [H, W, C], where H represents the height of the image, W represents the width of the image, and C represents the number of channels of the image. For conventional RGB images, C is usually equal to 3.

[0079] Convolutional layer: The convolutional layer is a key part of the neural network for feature extraction. In the YOLO network structure of the present invention, multiple layers of convolution operations are used, and each layer of convolution is followed by a Leaky ReLU activation function to increase the nonlinear expression ability of the model. The size of the convolution kernel is set to a combination of 3x3 and 1x1, which helps capture local details and global information in the image. The convolution step size is set to 1 to ensure that the spatial resolution of the image is not lost during the convolution process.

[0080] Feature fusion layer: In order to improve the model's ability to detect objects of different scales, the present invention uses a feature pyramid network structure to fuse feature maps of different scales. This layer can effectively combine information from different convolutional layers, thereby enhancing the model's recognition accuracy for multi-scale objects.

[0081] Prediction layer: In the prediction layer, the fused feature map is divided into multiple grids. Each grid is responsible for predicting a fixed number of bounding boxes and the corresponding category probabilities. In order to reduce overlap and inaccurate predictions, the system also uses non-maximum suppression (NMS) technology to screen the final detection results.

[0082] Output layer: The output layer is responsible for outputting the category, location, and confidence of the detected target. In addition, based on the continuous frame data, the system can also generate target trajectory sequences, which contain the target's continuous location information and timestamps, providing valuable data support for police management.

[0083] The model training steps are as follows:

[0084] Parameter initialization: Before starting training, you need to initialize the parameters of the YOLO network model, including the weights of the convolution kernel, bias terms, and learning rate. The reasonable setting of these parameters is crucial to the training speed and effect of the model.

[0085] Dataset partitioning: The preprocessed image dataset is partitioned into training, validation, and test sets. This partitioning helps evaluate the generalization ability of the model and ensures that the model performs well on unseen data.

[0086] Iterative training: The YOLO network model is iteratively trained using the training set. In each iteration, the model first performs forward propagation to calculate the output. Subsequently, the loss between the predicted value and the true value is calculated based on the loss function (including coordinate loss, confidence loss, and classification loss). The specific calculation formula of the loss function is: Loss = λ_coord Loss_coord + λ_noobj Loss_noobj + λ_obj Loss_obj + Loss_class, where the weights of each loss (λ_coord, λ_noobj, λ_obj) can be adjusted according to actual conditions.

[0087] Parameter update: Based on the calculated loss value, the model parameters are updated through the back propagation algorithm and the Adam optimizer. This process aims to minimize the loss function and thus improve the prediction accuracy of the model.

[0088] Training termination condition: Repeat the training steps until the loss function converges or reaches the preset number of iterations. The convergence criterion can be that the loss value does not change significantly in multiple consecutive iterations.

[0089] Model verification and testing: Use the verification set to verify the trained YOLO network model, mainly to evaluate the detection accuracy and recall rate of the model. When the model performance reaches the preset standard (such as both the accuracy and recall rate reach a certain threshold), the test set is used for the final test. If the test results meet the requirements, the model parameters are saved to obtain the trained target detection model.

[0090] Embodiment 2:

[0091] The behavior analysis model aims to identify and classify the abnormal behavior patterns of the target through in-depth analysis of the target trajectory sequence. The network structure of the behavior analysis model and the specific implementation methods of the training steps will be described in detail below.

[0092] The network structure includes:

[0093] Input layer: The input layer is responsible for receiving the target trajectory sequence output by the target detection model. Each sequence contains the continuous position information of the target and the corresponding timestamp. This information is the basic data for behavior analysis and provides rich spatiotemporal features for subsequent layers.

[0094] LSTM layer: The LSTM layer is the core of the behavior analysis model and consists of multiple layers of LSTM (Long Short-Term Memory) units. LSTM units are particularly suitable for processing sequence data and can effectively capture dependencies in the time dimension. In this model, the LSTM layer deeply processes the input target trajectory sequence and extracts temporal features related to the behavior pattern.

[0095] Fully connected layer: The fully connected layer is located after the LSTM layer. Its main function is to integrate and classify the time series features extracted by the LSTM layer. Through the processing of the fully connected layer, the model can identify different behavior patterns, especially abnormal behavior patterns, and provide key information for subsequent abnormal behavior detection.

[0096] Output layer: The output layer is the final output of the behavior analysis model, responsible for outputting the target’s abnormal behavior category and corresponding confidence level. These output results provide direct decision support for the police management system, helping to detect and handle abnormal situations in a timely manner.

[0097] The model training steps include:

[0098] Data preparation: The target trajectory sequences output by the target detection model are used as training sets. Among these sequences, known abnormal behavior categories are selected as labels for classification tasks in the supervised learning process.

[0099] Sequence segmentation: In order to adapt to the processing method of the LSTM model, the continuous target trajectories in the training set need to be segmented and divided into sequence segments of fixed length. This step helps the model better learn the time series features and improve the processing efficiency.

[0100] Iterative training: Use the divided sequence segments to iteratively train the LSTM network model. In each iteration, the model first performs forward propagation calculations to obtain predicted values. Then, the loss between the predicted value and the true value (i.e., the abnormal behavior category label) is calculated based on the cross entropy loss function.

[0101] Parameter update: Based on the calculated loss value, the model parameters are updated through the back propagation algorithm and the RMSprop optimizer. The RMSprop optimizer performs well in dealing with non-stationary targets and is suitable for parameter optimization in this model.

[0102] Training termination and verification: Repeat the training steps until the loss function converges or reaches the preset number of iterations. After that, use an independent verification set to verify the trained LSTM network model. Mainly evaluate the recognition accuracy and generalization ability of the model to ensure that the model can accurately identify abnormal behaviors in practical applications.

[0103] When the performance of the model on the validation set reaches the preset standard (such as the recognition accuracy exceeds a certain threshold), the model parameters are saved. In this way, a well-trained behavior analysis model is obtained, which can be deployed in the police management system for real-time monitoring and identification of abnormal behavior.

[0104] Embodiment 3:

[0105] The information integration module uses a rule engine to integrate the output results of the target detection model and the behavior analysis model to generate a police management report. The implementation of this module specifically includes:

[0106] Define events: Clarify the output format of the target detection model and the behavior analysis model. The output of the target detection model is defined as a detection event, each of which contains the target category (such as pedestrians, vehicles, etc.), location (coordinate information), time (detected timestamp), and confidence (the model's certainty about the detection). The output of the behavior analysis model is defined as a behavior event, each of which contains the category of the behavior (such as abnormal aggregation, rapid movement, etc.), target ID (used to identify a specific target), time period (the time range in which the behavior occurs), and confidence.

[0107] According to the preset rules, the detection events and behavior events are analyzed for correlation. The purpose of this step is to find the potential connection between the two, so as to more fully understand the dynamics and behavior of the target. Specifically include:

[0108] The association condition is set based on the information in the detection event and the behavior event. For example, the association condition can be set as follows: the target ID in the detection event is the same as the target ID in the behavior event, and the timestamps or time periods of the two overlap. In addition, factors such as location information, target category, and behavior category can also be considered to enhance the accuracy and effectiveness of the association.

[0109] Match the detection events and behavior events that meet the association conditions to form associated events. This step can be achieved through database query, algorithm matching, etc. After the match is successful, a list of associated events containing richer information will be obtained.

[0110] Conduct risk assessment on related events to determine the risk level of each event. Risk assessment can be based on multiple dimensions such as the nature of the event, the time and location of occurrence, and the target's behavior pattern. For example, abnormal gathering behavior in certain specific areas or time periods may be considered a high-risk event.

[0111] Generate a police management report based on the associated events and risk levels. This step is designed to provide decision support and help police managers respond to and handle abnormal situations in a timely manner. Specifically include:

[0112] Select recommended measures: Select corresponding recommended measures from the preset recommended measures library according to the associated events and risk levels. These recommended measures may include strengthening patrols, investigating suspicious targets, issuing safety warnings, etc. The selection of recommended measures should be based on actual conditions and historical experience to ensure their effectiveness and feasibility.

[0113] Combined report content: Combine related events, risk levels and selected recommended measures into a police management report. The report should present key information in a clear and concise manner to facilitate police managers to quickly understand and make decisions.

[0114] Output report: Output police management reports for police management personnel to make reference for decision-making. Reports can be sent to relevant personnel via email, SMS, mobile applications, etc. to ensure timely delivery and response of information.

[0115] The present invention also includes a police management method based on image recognition technology, the method comprising:

[0116] Step 1: Receive video stream and image data from the surveillance camera;

[0117] Step 2: Perform format conversion, denoising, enhancement and feature extraction on the received data to form a preprocessed image data set;

[0118] Step 3: Build a target detection model based on a deep convolutional neural network algorithm and use the preprocessed image dataset for training to detect specific targets in the image and output a trajectory sequence of the target. Each sequence contains the continuous position information and timestamp of the target.

[0119] Step 4: Build a behavior analysis model based on the long short-term memory network algorithm, and use the target trajectory sequence output by the target detection model to analyze and identify the abnormal behavior pattern of the target;

[0120] Step 5: Integrate the output results of the target detection model and the behavior analysis model to generate a police management report.

[0121] The implementation of this method refers to the above embodiment and will not be described in detail in the specification.

[0122] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0123] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A police management system based on image recognition technology, characterized in that: The system comprises: A data receiving module, used for receiving video stream and image data from a surveillance camera; The data preprocessing module performs format conversion, denoising, enhancement and feature extraction on the received data to form a preprocessed image data set; The target detection module builds a target detection model based on a deep convolutional neural network algorithm. The model is trained using a preprocessed image dataset to detect specific targets in the image and output a trajectory sequence of the target. Each sequence contains the continuous position information and timestamp of the target. The behavior analysis module builds a behavior analysis model based on the long short-term memory network algorithm. The model uses the target trajectory sequence output by the target detection model to analyze and identify the abnormal behavior pattern of the target; The information integration module integrates the output results of the target detection model and the behavior analysis model to generate a police management report; The information integration module uses a rule engine to integrate the output results of the target detection model and the behavior analysis model. The specific steps include: The output of the target detection model is defined as a detection event, which includes the target category, location, time and confidence. The output of the behavior analysis model is defined as a behavior event, which includes the behavior category, target ID, time period, and confidence level. According to the preset rules, the detection events and behavior events are correlated and analyzed; Generate police management reports containing relevant incidents, risk levels and recommended actions.

2. According to claim 1, a police management system based on image recognition technology is characterized in that: The target detection model adopts the YOLO network structure, and its network structure includes: Input layer: receives preprocessed image data, and the image size is unified to [H, W, C], where H is the height, W is the width, and C is the number of channels; Convolutional layer: multi-layer convolution operation is adopted, each convolution layer is followed by a Leaky ReLU activation function, the convolution kernel size is set to a combination of 3x3 and 1x1, and the step size is set to 1; Feature fusion layer: fuses feature maps of different scales through the feature pyramid network structure; Prediction layer: The fused feature map is divided into grids, each grid predicts a fixed number of bounding boxes and category probabilities, and non-maximum suppression is used to filter the final detection results; Output layer: Outputs the category, location, and confidence of the detected target, as well as the target trajectory sequence generated based on continuous frame data, which contains the continuous location information and timestamp of the target.

3. The police management system based on image recognition technology according to claim 2 is characterized in that: The training steps of the target detection model include: Initialize the parameters of the YOLO network model, including convolution kernel weights, bias terms, and learning rates; Divide the preprocessed image dataset into training set, validation set and test set; The YOLO network model is iteratively trained using the training set. In each iteration, the forward propagation calculates the model output, and the loss between the predicted value and the true value is calculated according to the loss function. The loss function includes coordinate loss, confidence loss, and classification loss. The calculation formula is: Loss = λ_coord Loss_coord + λ_noobj Loss_noobj + λ_obj Loss_obj +Loss_class Among them, Loss_coord is the coordinate loss, Loss_noobj is the confidence loss of the non-target box, Loss_obj is the confidence loss of the target box, and Loss_class is the classification loss; λ_coord, λ_noobj, λ_obj and Loss_class are the weights of Loss_coord, Loss_noobj, Loss_obj and Loss_class respectively; According to the loss value, the model parameters are updated through the back-propagation algorithm and the Adam optimizer; Repeat the training steps until the loss converges or the preset number of iterations is reached; The trained YOLO network model is verified using the validation set to evaluate the detection accuracy and recall rate of the model. When the model performance reaches the preset standard, the test set is used for the final test, and the model parameters are saved to obtain the trained target detection model.

4. The police management system based on image recognition technology according to claim 1 is characterized in that: The network structure of the behavior analysis model includes: Input layer: receives the target trajectory sequence output by the target detection model. Each sequence contains the continuous position information and timestamp of the target. LSTM layer: contains multiple layers of LSTM units. Each layer of LSTM units processes the input sequence and captures the dependencies in the time dimension. Fully connected layer: The output features of the LSTM layer are classified through the fully connected layer to identify the abnormal behavior patterns of the target; Output layer: Outputs the abnormal behavior category and confidence of the target.

5. The police management system based on image recognition technology according to claim 4 is characterized in that: The steps to train a behavior analysis model include: Use the target trajectory sequence output by the target detection model as the training set and select the abnormal behavior category as the label; Sequence the training set and divide the continuous target trajectory into sequence segments of fixed length; The LSTM network model is iteratively trained using sequence segments. In each iteration, the model output is calculated by forward propagation, and the loss between the predicted value and the true value is calculated according to the cross entropy loss function. According to the loss value, the model parameters are updated through the back-propagation algorithm and the RMSprop optimizer; Repeat the training steps until the loss converges or the preset number of iterations is reached; The trained LSTM network model is verified using the validation set to evaluate the recognition accuracy and generalization ability of the model. When the model performance reaches the preset standard, the model parameters are saved to obtain the trained behavior analysis model.

6. The police management system based on image recognition technology according to claim 1 is characterized in that: The step of performing correlation analysis on the detection event and the behavior event according to the preset rules further includes: Setting an association condition, wherein the association condition is based on at least one of the target category, location, time, and confidence in the detection event, and the behavior category, target ID, time period, and confidence in the behavior event; Matching detection events and behavior events that meet the association condition to form an associated event; Conduct a risk assessment on the associated events and determine the risk level based on the risk assessment results.

7. The police management system based on image recognition technology according to claim 6 is characterized in that: The step of generating a police management report including associated events, risk levels and recommended actions further includes: According to the associated events and risk levels, select corresponding recommended measures from the preset recommended measures library; Combining the associated events, risk levels and selected recommended actions into a police management report; The police management report is output for reference by police management personnel in making decisions.

8. The police management system based on image recognition technology according to claim 1 is characterized in that: The data preprocessing module uses the principal component analysis (PCA) algorithm to extract features, and the specific steps include: Calculate the covariance matrix of image data; Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues; Select the eigenvectors corresponding to the first k largest eigenvalues ​​to form the feature space; The original image data is projected into the feature space to obtain the reduced-dimensional feature data as the model input.

9. A police management method based on image recognition technology, characterized in that: The method comprises: Step 1: Receive video stream and image data from the surveillance camera; Step 2: Perform format conversion, denoising, enhancement and feature extraction on the received data to form a preprocessed image data set; Step 3: Build a target detection model based on a deep convolutional neural network algorithm and use the preprocessed image dataset for training to detect specific targets in the image and output a trajectory sequence of the target. Each sequence contains the continuous position information and timestamp of the target. Step 4: Build a behavior analysis model based on the long short-term memory network algorithm, and use the target trajectory sequence output by the target detection model to analyze and identify the abnormal behavior pattern of the target; Step 5: Integrate the output results of the target detection model and the behavior analysis model to generate a police management report; the specific steps include: The output of the target detection model is defined as a detection event, which includes the target category, location, time and confidence. The output of the behavior analysis model is defined as a behavior event, which includes the behavior category, target ID, time period, and confidence level. According to the preset rules, the detection events and behavior events are correlated and analyzed; Generate police management reports containing relevant incidents, risk levels and recommended actions.

Images