Intelligent camera small target object tracking identification method

Abstract

The invention discloses an intelligent camera small target object tracking identification method, which belongs to the technical field of target identification, and specifically comprises the following steps: collecting small target and standard target data through a zoom camera, and marking the type; and according to different types of small target data, generating a small target detection algorithm model based on a YOLOv8 detection model. Field images are collected in real time and preprocessed, and a small target detection algorithm model is used for recognizing the type of a small target. Obtaining a set amplification threshold value under a camera coordinate system according to the type of the small target, and calculating the size of a detection frame of the small target. If the size of the detection frame is greater than or equal to the set amplification threshold, violation detection is directly carried out; if the size is smaller than the set amplification threshold, adjusting the focal length and angle of the camera, carrying out amplification by taking the central point of the detection frame area as a reference, and then carrying out violation detection on the small target; according to the invention, the accuracy and efficiency of small target detection are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target recognition, and in particular to an intelligent camera small target object tracking and recognition method. Background Art

[0002] With the continuous advancement of deep learning technology, rapid prototyping technology has also been increasingly widely used in the field of image recognition. This technology can quickly learn features from a large amount of data through algorithms and apply them to new data sets for accurate classification or detection. However, in actual applications, especially in surveillance scenarios, the recognition of small target objects still faces many challenges.

[0003] In order to improve the accuracy of small object recognition, researchers proposed a method based on image segmentation. The core idea of ​​this method is to divide the original image into multiple small blocks according to certain rules, and then process and analyze these small blocks separately. For large-size images with high resolution and rich details, this method can effectively improve the model's attention to local features, thereby improving overall performance.

[0004] When faced with images captured by low-pixel cameras, it is often difficult to achieve the desired effect by simply increasing the number of training samples or adjusting the network structure. This is because the image itself contains limited information, and even after fine segmentation, it may not provide enough clues for machine learning. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent camera small target object tracking and recognition method to solve the following technical problems: When faced with images captured by a low-pixel camera, it is often difficult to achieve the desired effect simply by increasing the number of training samples or adjusting the network structure.

[0006] The purpose of the present invention can be achieved through the following technical solutions: An intelligent camera small target object tracking and recognition method comprises the following steps: Collect target image data through a variable-focus camera, the target image data including small target data and standard target data, and mark the target image data with normal behavior type or illegal behavior type, and generate corresponding small target detection algorithm models according to different types of target data; The real-time image of the scene is collected by a zoom camera, and the real-time image is preprocessed. The small target area in the real-time image is identified by a small target detection algorithm model, and it is determined whether the small target area is a violation. The illegal small target is marked as a violation target, and the set magnification threshold in the camera coordinate system is obtained according to the type of the violation target, and the detection frame size of the violation target in the camera coordinate system is calculated; the detection frame size is compared with the set magnification threshold; If the size of the detection frame is greater than or equal to the set enlargement threshold, the judgment result of the illegal target will be directly reported; If the size of the detection frame is smaller than the set magnification threshold, the current real-time image is discarded, the camera focal length and angle are adjusted according to the coordinates of the detection frame, and the detection frame area is enlarged based on the center point. The illegal target is judged again based on the enlarged real-time image.

[0007] As a further solution of the present invention: the small target data represents the target image in the scene of distant view and panoramic view, and the standard target data represents the target image in the scene of medium view, close view and close-up.

[0008] As a further solution of the present invention: use the labelimg image annotation tool to perform corresponding types of annotations on the collected small target data and standard target data.

[0009] As a further solution of the present invention: the normal behavior types include behaviors that comply with construction safety regulations on the construction site; the illegal behavior types include behaviors that do not comply with construction safety regulations on the construction site and personal dangerous actions.

[0010] As a further solution of the present invention: the process of preprocessing the real-time image is: The real-time image is scaled and the pixels of the real-time image are interpolated or merged to make the real-time image conform to the input size set by the model. The preprocessed image is input into the model for inference detection. The detection results include the target box coordinates, confidence, and category.

[0011] As a further solution of the present invention, the specific process of enlarging based on the center point of the detection frame area is as follows: Get the magnification A of this type of small target and calculate the length C that the Huawei camera should capture at this magnification x , and the width C that the Huawei camera should capture at this magnification y , the formula is: ; Among them, 355 is the maximum parameter of the selected zoom camera fixed as the x-axis, and 288 is the maximum parameter of the selected zoom camera fixed as the y-axis. The calculation formula for the upper left corner coordinates (X1, Y1) and the lower right corner coordinates (X2, Y2) of the detection frame after the small target is enlarged is: ; ; Among them, X c is the x-axis coordinate of the center point of the detection frame, Y c is the y-axis coordinate of the center point of the detection frame, W is the width of the real-time image, and H is the height of the real-time image.

[0012] As a further solution of the present invention: the process of adjusting the focal length and angle of the camera according to the coordinates of the detection frame includes: The camera control engineering code is deployed to the algorithm detection server, including the camera login interface, camera zoom and magnification interface, and camera return to preset position interface. If the detection frame size is smaller than the set magnification threshold, the login interface is called to log in to the camera. After successful login, the zoom interface and magnification interface are called to zoom in on the target center. The real-time image is captured again and returned to the algorithm detection server. The camera preset position parameter is changed to 1. After waiting for the set time, the return to preset position interface is called to adjust the focal length and angle to the original position of the camera, and the camera preset position parameter is changed to 0.

[0013] Beneficial effects of the present invention: The present invention uses a small target detection algorithm model based on the YOLOv8 detection model, and the method can more accurately identify and track small target objects; if the size of the detected small target is smaller than the set magnification threshold, the system will automatically adjust the focal length and angle of the camera to magnify the target area, thereby improving the recognition rate of the small target. This adaptive magnification mechanism ensures that small targets can be effectively detected even at low resolution or long distances. For blurred small targets, the false detection caused by unclear small targets and few features is reduced. Automatic tracking of targets is achieved, and violations are confirmed multiple times to reduce the situation of false alarms. The zoom magnification function of the camera has a significant improvement effect on the detection of complex scenes such as backlighting and blur, and can further improve the algorithm detection accuracy. By combining advanced deep learning technology, adaptive magnification mechanism and flexible processing flow, the present invention significantly improves the accuracy and efficiency of small target detection, while reducing the false alarm rate, providing strong technical support for monitoring and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The present invention will be further described below in conjunction with the accompanying drawings.

[0015] Figure 1 is a flowchart of the present invention based on the installation and deployment of smart camera devices and systems to algorithm reasoning; Figure 2 is a flow chart of the present invention for performing zoom control on a camera based on the algorithm detection result; Figure 3 It is a communication flow chart of the camera, algorithm and platform of the present invention. DETAILED DESCRIPTION

[0016] 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.

[0017] See also Figure 1-Figure 3 As shown, the present invention is an intelligent camera small target object tracking and recognition method, comprising the following steps: a. The present invention adopts a variable-focus binocular camera to capture pictures by controlling the change of camera focal length and angle rotation, which is used to obtain small target training data and use the detection algorithm model to infer pictures in monitoring scenarios.

[0018] b. Training small target detection algorithm model. In the present invention, the yolov8 detection model is used, and the small target data is collected in a targeted manner, the small target data is annotated, the model parameters are optimized, the training optimizer is set to SGD, the input image size is set to 1024*1024, the learning rate is set to 0.0003, the training step size is set to 200 generations, and the annotated categories that need to be enlarged are trained until the loss function no longer decreases, and the best model is selected as the algorithm detection model of the present invention.

[0019] c. Set up multiple patrol and watch positions on site to monitor the large scene on site. The camera is equipped with multiple detection algorithms to monitor multiple illegal behaviors on the site at the same time and detect illegal behaviors at the same time. When there are multiple illegal behaviors in the picture, they will be zoomed in and confirmed in sequence.

[0020] d. The algorithm detects that the server device communicates with the camera to infer the image sent back from the image acquisition module, and controls the focal length and angle of the corresponding camera to zoom in on the target and re-infer according to the inference result.

[0021] e. Alarm platform: Send zoom instructions to the camera based on the algorithm inference results, and display the alarm results.

[0022] The specific process of algorithm intelligent tracking: 1. Install intelligent zoom cameras at the production site, and set up multiple patrol and watch positions on the on-site intelligent cameras to monitor on-site violations; 2. The camera is associated with the algorithm detection server so that the algorithm detection server can obtain the video stream of the zoom camera and control the focal length and rotation angle of the smart zoom camera.

[0023] 3. Algorithm model reasoning. First, the algorithm server preprocesses the image. The algorithms for different scene layouts are processed differently. The target detection algorithm scales the image before inputting it into the algorithm model to ensure that the input image meets the input size set by the algorithm model. The preprocessed image is input into the model for reasoning detection. The detection result includes information such as target box coordinates, confidence, and category.

[0024] 4. Set the zoom magnification thresholds for different algorithms. First, set the zoom parameter thresholds for the algorithm to detect the target frame: different thresholds can be configured according to different algorithms. For example, at a picture resolution of 1920x1080, the zoom magnification threshold for the glove detection target frame is set to 24x24, and the magnification threshold for the mask is set to 32x32. If the size of the alarm category target frame is larger than the above set value, we believe that the target is large and has a high credibility and can be reported directly. If the size of the identified target frame is smaller than the above value, we believe that the category needs to be confirmed again. The parameters we set are based on a resolution of 1920x1080. If the picture is of other sizes, the zoom parameter threshold for the algorithm to detect the target frame will be enlarged in proportion. For example, when the picture size is 2560x1440, the magnification threshold for the mask is 42x42.

[0025] A refers to the set enlargement threshold. In the above example, A mask is 32, which refers to the first parameter in the enlargement threshold of other image sizes, which refers to the second parameter in the enlargement threshold, W refers to the width of the image, and H refers to the height of the image. In this way, even if the size of the input image is different, the enlargement threshold can be set adaptively according to the width and height of the image.

[0026] 5. Calculate the camera magnification and coordinates based on the target detection results. Including: Calculate the width and height of the target detection frame in the image, and calculate the area of ​​the target frame. If the area of ​​the target frame coordinates is larger than the area of ​​the set magnification threshold, the target is considered to be a large target and can be identified. For large targets, no magnification is required and the result can be reported directly. If the target frame area is smaller than the threshold, the target is considered to be a small target and the camera needs to be zoomed in for secondary confirmation, and the result of this recognition is discarded, waiting for the camera to send the image again for recognition after the zoom operation. In the inspection task, the camera only zooms in once for the algorithm recognition, and the magnification ratio needs to be set in advance. For example, the magnification ratio of gloves can be set to 3 times to confirm that gloves can be recognized at this ratio.

[0027] 6. After the second confirmation of the violation, wait for 20 seconds, adjust the camera back to the watching position, continue to identify the next inspection task, monitor the violation of the entire large scene, and repeat steps 1-6. In order to avoid endless zooming of the camera after the algorithm detects the target, the present invention stipulates that the camera only zooms in once for an algorithm. When in the watching position, the preset position parameter of the camera is 0. When the camera tracks and zooms in on the algorithm, the preset position parameter is changed to 1. Then it is necessary to wait for 20 seconds for the camera to be adjusted back to the watching position, and the preset position parameter is changed to 0 to continue monitoring the entire large scene.

[0028] In another preferred embodiment of the present invention: like Figure 1 As shown, first of all, in the workplaces of personnel in the implementation site of this project, Huawei X series binocular zoom cameras are installed facing the construction direction to cover most of the workplaces.

[0029] Install the algorithm inference equipment, including: prepare the algorithm server, the algorithm inference server has 4*TeslaT4 graphics cards, network connection equipment. Install the graphics card driver corresponding to the graphics card version, cuda graphics card acceleration package and dependent libraries required to run the algorithm in the server, receive the pictures captured by the camera and input them into the inspection algorithm model for recognition and detection.

[0030] After installing the camera, collect small target data with small image size such as smoking, helmets, heads, safety belts, people falling to the ground, etc. in large scenes, as well as normal data with large image size such as people, helmets, heads, people falling to the ground, etc. in normal scenes. Use the labelimg image annotation tool to annotate the collected small target data and normal data with corresponding categories. In the category definition of smoking and helmets, putting a cigarette to the mouth is defined as the smoking category, and holding a cigarette in the hand or clamping it to the ear does not belong to the smoking category. Common helmets on the market, the combination of helmets and straw hats, the combination of helmets and sun hats, the combination of helmets and warm hats, etc., all belong to the helmet category. People who do not wear any covering on their heads belong to the category of not wearing helmets.

[0031] Training algorithm model. Using yolov8 as the basic model of the target detection algorithm of the present invention, the training optimizer is set to SGD, the input image size is set to 1024*1024, the learning rate is set to 0.0003, the training step size is set to 200 generations, and the best model is selected as the algorithm detection model of the present invention.

[0032] Deploy the algorithm. Deploy the algorithm model to the inference server, including: setting the size of the captured image, setting the network transmission protocol, configuring the algorithm environment, setting the algorithm output results, debugging the algorithm model and testing it.

[0033] Deploy the camera control project. Deploy the camera control project code to the algorithm detection server, including: camera login interface, camera zoom interface, camera return preset position interface. The camera focal length is controlled through these three interfaces to achieve automatic recognition and tracking of the target.

[0034] Test the detection effect. Specifically include: test the detection effect of the algorithm, test the camera interface, test the communication between the equipment, and test the display interface of the alarm platform.

[0035] In another preferred embodiment of the present invention: like Figure 2 As shown in the example, the method described is to analyze the algorithm model from the captured image of the camera and control the camera to track the target. The method includes: 1. The camera captures pictures of the construction scene and transmits the captured pictures back to the algorithm analysis server. In this embodiment, the pixel size of the camera is 2560x1440.

[0036] 2. The algorithm server pre-processes the image, and the algorithms for different scene layouts perform different processing, including: The helmet and person recognition detection algorithm resizes the image before inputting it into the algorithm model to ensure that the input image meets the input size set by the algorithm model. The smoking algorithm cuts the image into four equal parts and then resizes each image to make each image meet the input size of the algorithm model.

[0037] 3. Algorithm model reasoning. Input the preprocessed images into the model for reasoning and detection, including: The images of helmets and people are inferred and tested. The test results include the target frame coordinates, confidence, and category. The cut images of smoking are tested in batches. The test results are transformed to output the coordinate points of a rectangular frame and plotted on the coordinates of the corresponding position of the original image. Finally, the tested images are merged into an image of the same size as the original image.

[0038] Target result optimization. The algorithm model detection results are sent to the optimization center, which includes: An optional confidence filtering function can be set to select the confidence of different algorithms according to different scenarios, and the results with less than the set confidence will be filtered out, which can more flexibly meet the needs of different scenarios.

[0039] Set the optional personnel filtering function. The helmet algorithm, smoking algorithm, seat belt algorithm, etc. can be added to the personnel filtering. When the helmet, smoking, and seat belt are detected, the personnel detection algorithm is used to detect the personnel, and it is determined whether there is an intersection between the detection rectangle of the helmet and the detection rectangle of the person. The detection results without intersection will be filtered to reduce false alarms caused by the background.

[0040] Set an optional area frame. The safety belt algorithm needs to demarcate the high-altitude working area. People outside the area do not need to wear safety belts. The personnel leaving the post algorithm needs to demarcate the work area. People who leave their work posts and do not return within ten minutes are judged as leaving their posts.

[0041] 5. Target frame enlargement calculation, including: Set the optional suspicion value size. It is generally calculated based on the average size of the target frame of the algorithm category. When the target detection frame size returned by the algorithm is larger than the suspicion value, it does not belong to a small target, and the violation is directly reported to the platform. When the target detection frame returned by the algorithm is within the suspicion value, the current detection result is discarded, and the coordinates of the detection result are recorded and converted into the focal length and rotation angle of the camera. The camera zooms in on the target as the center point.

[0042] 6. According to the results, call the camera interface to track and amplify the detection results. Send the above calculation results and parameters to the camera. After the detection results are optimized by the optimization center, the target size will be judged. If the target size is within the suspected value range, call the login interface to log in to the camera. After successful login, call the zoom and zoom interface to zoom in on the target center, and capture the picture again and return it to the algorithm detection server. After 20 seconds, call the return to preset position interface to adjust the focal length to the original position of the camera.

[0043] 7. Get the image again for inference and send it to the alarm platform. The algorithm server receives the enlarged image for inference and processes it again through the above optimization center. If there is any violation in the detection result, it will be reported to the alarm platform for display.

[0044] In another preferred embodiment of the present invention: like Figure 3 As shown, the case describes the communication process between the camera, the algorithm and the platform. The method includes: 1. The camera at the construction site captures pictures at a fixed time and sends them to the detection equipment.

[0045] 2. After the detection device receives the image, it first performs pre-processing on the image by reshaping the image size, cutting the image, etc., and then inputs the image into the algorithm model for inference. The inference result is optimized by the optimization center in the detection device and the result is sent to the alarm platform.

[0046] 3. After receiving the inference results, the alarm platform transmits the zoom command back to the camera. The camera executes the zoom image command and re-captures the image and transmits it to the detection equipment. After being processed by the pre-processing and optimization center again, the inference results are transmitted back to the alarm platform. If any violation occurs, it will be displayed on the alarm platform.

[0047] In another preferred embodiment of the present invention: The case described is: the network structure of the target detection box.

[0048] The main structure of yolov8 is basically composed of backbone, neck and head. Backbone is the trunk network of YOLOv8, which is mainly used to extract the features of the input image. With reference to the CSPDarkNet-53 network, the C2f (CSPLayer_2Conv) structure is used to replace the C3 module in YOLOv5. The C2f structure cancels the convolution operation in the branch while adding skip layer connections, and adds additional split operations, thereby enriching the feature information and reducing the amount of calculation. In the Neck part, YOLOv8 adds a SPPF module, which can pool features at different scales to capture targets of different sizes. Head is the detection layer of YOLOv8, and Decoupled Head is used in YOLOv8, which is different from the Coupled Head used in YOLOv3, YOLOv4, and YOLOv5. YOLOv8 also uses 3 output branches, each of which is divided into 2 parts, which are used for classification and regression borders respectively.

[0049] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0050] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0051] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0052] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0053] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0054] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0055] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0056] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0057] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0058] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent camera small target object tracking and recognition method, characterized in that: The following steps are involved: Collect target image data through a variable-focus camera, the target image data including small target data and standard target data, and mark the target image data with normal behavior type or illegal behavior type, and generate corresponding small target detection algorithm models according to different types of target data; The real-time image of the scene is collected by a zoom camera, and the real-time image is preprocessed. The small target area in the real-time image is identified by a small target detection algorithm model, and it is determined whether the small target area is a violation. The illegal small target is marked as a violation target, and the set magnification threshold in the camera coordinate system is obtained according to the type of the violation target, and the detection frame size of the violation target in the camera coordinate system is calculated; Compare the detection box size with the set magnification threshold; If the size of the detection frame is greater than or equal to the set enlargement threshold, the judgment result of the illegal target will be directly reported; If the size of the detection frame is smaller than the set magnification threshold, the current real-time image is discarded, the camera focal length and angle are adjusted according to the coordinates of the detection frame, and the detection frame area is enlarged based on the center point. The illegal target is judged again based on the enlarged real-time image.

2. The intelligent camera small target object tracking and recognition method according to claim 1 is characterized in that: The small target data represents the target image in the scene with distant view and panoramic view, and the standard target data represents the target image in the scene with medium view, close view and close-up view.

3. The intelligent camera small target object tracking and recognition method according to claim 2 is characterized in that: Use the labelimg image annotation tool to annotate the collected small target data and standard target data with corresponding types.

4. The intelligent camera small target object tracking and recognition method according to claim 1 is characterized in that: The normal behavior types include behaviors that comply with construction site safety regulations; the illegal behavior types include behaviors that do not comply with construction site safety regulations and personal danger actions.

5. The intelligent camera small target object tracking and recognition method according to claim 1 is characterized in that: The process of preprocessing real-time images is as follows: The real-time image is scaled and the pixels of the real-time image are interpolated or merged to make the real-time image conform to the input size set by the model. The preprocessed image is input into the model for inference detection. The detection results include the target box coordinates, confidence, and category.

6. The intelligent camera small target object tracking and recognition method according to claim 1, characterized in that: The specific process of enlarging based on the center point of the detection frame area is as follows: Get the magnification A of this type of small target and calculate the length C that the Huawei camera should capture at this magnification x , and the width C that the Huawei camera should capture at this magnification y , the formula is: ； Among them, 355 is the maximum parameter of the selected zoom camera fixed as the x-axis, and 288 is the maximum parameter of the selected zoom camera fixed as the y-axis. The calculation formula for the upper left corner coordinates (X1, Y1) and the lower right corner coordinates (X2, Y2) of the detection frame after the small target is enlarged is: ； ； Among them, X c is the x-axis coordinate of the center point of the detection frame, Y c is the y-axis coordinate of the center point of the detection frame, W is the width of the real-time image, and H is the height of the real-time image.

7. The intelligent camera small target object tracking and recognition method according to claim 1 is characterized in that: The process of adjusting the camera focal length and angle according to the coordinates of the detection frame includes: The camera control engineering code is deployed to the algorithm detection server, including the camera login interface, camera zoom and magnification interface, and camera return to preset position interface. If the detection frame size is smaller than the set magnification threshold, the login interface is called to log in to the camera. After successful login, the zoom interface and magnification interface are called to zoom in on the target center. The real-time image is captured again and returned to the algorithm detection server. The camera preset position parameter is changed to 1. After waiting for the set time, the return to preset position interface is called to adjust the focal length and angle to the original position of the camera, and the camera preset position parameter is changed to 0.

Images