Camera dynamic visual angle adjustment method and device based on deep learning, network camera and storage medium

Abstract

The invention relates to the technical field of intelligent control, in particular to a camera dynamic visual angle adjusting method and device based on deep learning, a network camera and a storage medium, and the method comprises the steps: obtaining a current working scene type and a target detection demand of the camera; preliminarily estimating a basic view angle coverage range based on a scene type, calculating a view angle adjustment coefficient in combination with a target detection demand, initializing a camera view angle control parameter, and collecting current scene features in real time through a scene sensor; establishing a deep learning dynamic optimization model, taking scene features and initial parameters as input, analyzing a change trend of a visual angle coverage effect, generating an optimal visual angle control parameter through iterative optimization, and ensuring accurate matching of a visual angle range and a target detection demand in a complex scene; the camera operation parameters are adjusted in real time based on the optimal parameters, dynamic adaptive control of the visual angle is achieved, the integrity and accuracy of target detection in a monitoring scene are improved, and the problem that a traditional fixed visual angle scheme is insufficient in coverage is solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and in particular to a method and device for dynamic viewing angle adjustment of a camera based on deep learning, a network camera, and a storage medium. Background Art

[0002] With the widespread application of technologies such as intelligent monitoring, security systems, and human-computer interaction, there is a growing demand for dynamic adjustment of camera perspectives in different scenarios. Traditional camera perspective control methods often rely on fixed preset parameters or manual adjustments, making them difficult to adapt to complex and changing work scenarios. In existing technologies, perspective adjustment strategies often ignore the coupling relationship between scene characteristics and target detection requirements, resulting in blind spots in perspective coverage, reduced target recognition accuracy, or waste of resources. For example, in high-density crowd scenes, a fixed perspective may not be able to take into account both global monitoring and key target tracking; in dynamic lighting environments, a single perspective parameter makes it difficult to balance image clarity and coverage.

[0003] The above content is only used to assist in understanding the technical solution of the present invention and does not constitute an admission that the above content is prior art. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method, device, network camera, and storage medium for dynamic camera perspective adjustment based on deep learning, aiming to solve the technical problems in the existing technology that the camera perspective adjustment strategy often ignores the coupling relationship between scene characteristics and target detection requirements, resulting in blind spots in perspective coverage, reduced target recognition accuracy or waste of resources.

[0005] To achieve the above objectives, the present invention provides a method for dynamic camera viewing angle adjustment based on deep learning, the method comprising: Get the camera's current working scene type and target detection requirements; Preliminarily estimating the camera's viewing angle coverage according to the current working scene type, and determining the camera's viewing angle adjustment coefficient based on the target detection requirements and the preliminarily estimated viewing angle coverage; Initializing the camera's viewing angle control parameters based on the camera's viewing angle adjustment coefficient, and determining current scene features based on preset scene sensor recognition; Evaluating the camera's viewing angle coverage effect based on the current scene features and the viewing angle control parameters of the initialized camera; Establishing a dynamic view control optimization model based on deep learning, analyzing the change trend of the view coverage effect of the camera based on the current scene characteristics and the view control parameters of the initialized camera, and generating the optimal view control parameters; The viewing angle operating parameters of the camera are optimized and controlled based on the optimal viewing angle control parameters.

[0006] Optionally, before initializing the camera's view angle control parameters based on the camera's view angle adjustment coefficient and determining the current scene features based on preset scene sensor recognition, the method further includes: Obtaining a working scene type of the camera and several scene features of the working scene type, and determining a viewing angle control range corresponding to the scene features of the working scene type, wherein the scene features of the working scene type include scene complexity, target motion speed, and target size; Accordingly, the camera viewing angle adjustment coefficient is based on the camera, the viewing angle control parameter of the camera is initialized, and the current scene feature is determined based on the preset scene sensor recognition, including: The viewing angle adjustment algorithm is used to correct the viewing angle control range corresponding to the scene characteristics of the working scene type based on the viewing angle adjustment coefficient of the camera, determine the viewing angle control parameters of the initialized camera, and determine the current scene characteristics based on the preset scene sensor recognition.

[0007] Optionally, the viewing angle adjustment algorithm is expressed as:

[0008] Where, To initialize the camera's viewing angle control parameters, is the maximum viewing angle control range corresponding to the scene characteristics of the work scene type, is the minimum viewing angle control range corresponding to the scene characteristics of the working scene type, and E is the viewing angle adjustment coefficient of the camera.

[0009] Optionally, the evaluating the camera's viewing angle coverage effect according to the current scene feature and the initialization camera's viewing angle control parameter includes: Assigning a coverage risk score to each scene feature according to a coverage risk change trend that occurs when several scene features in the current scene feature affect camera control, thereby obtaining a scene feature coverage risk score; Standardizing the scenario feature coverage risk score through a standardized formula; Based on the standardized scene feature coverage risk scores, determine the relative importance of each scene feature coverage risk score to the view coverage during camera control, assign a coverage risk weight to each scene feature, and obtain the scene feature coverage risk weight; Calculating a camera's coverage weighted score based on the scene feature coverage risk score and the scene feature coverage risk weight; Based on the coverage weighted score of the camera and the initialization camera's viewing angle control parameters, the camera's viewing angle coverage effect is calculated through a coverage evaluation formula.

[0010] Optionally, the coverage evaluation formula is:

[0011] Where C is the camera's viewing angle coverage effect, is the risk weight of the camera’s i-th scene feature coverage, is the coverage risk score of the i-th scene feature of the camera, V is the initialization camera's viewing angle control parameter, and n is the total number of scene features of the camera.

[0012] Optionally, the establishing of a dynamic viewing angle control optimization model based on deep learning, analyzing the changing trend of the viewing angle coverage effect of the camera based on the current scene features and the viewing angle control parameters of the initialized camera, and generating the optimal viewing angle control parameters, includes: The viewing angle control range corresponding to the scene characteristics of the work scene type is used as an adjustable viewing angle control restriction condition; Construct a dynamic view control optimization model based on deep learning; Based on the deep learning-based dynamic perspective control optimization model, under the adjustable perspective control restriction condition, the current scene features and the perspective control parameters of the initialized camera are used as independent variable inputs to determine the changing trend of the perspective coverage effect of the camera. The perspective control parameters are gradually increased according to the adjustable perspective control restriction condition, and several rounds of iterations are performed to output the perspective control parameters that maximize the perspective coverage effect of the camera as the dependent variable, thereby generating the optimal perspective control parameters.

[0013] Optionally, the expression of the dynamic viewing angle control optimization model is:

[0014] Where, is the optimal viewing angle control parameter, For the The viewing angle control parameters during round iteration, is the learning efficiency step size, is the gradient of the viewing angle control parameter, is the camera's view coverage effect function, V is the initialization view control parameter, S is the current scene feature, is the upper limit of the viewing angle control range corresponding to the scene characteristics of the work scene type, is the lower limit.

[0015] In addition, to achieve the above-mentioned purpose, the present invention further provides a camera dynamic viewing angle adjustment device based on deep learning, the camera dynamic viewing angle adjustment device based on deep learning comprising: The scene acquisition module is used to obtain the camera's current working scene type and target detection requirements; A perspective preliminary estimation module, configured to preliminarily estimate the perspective coverage of the camera according to the current working scene type, and determine the perspective adjustment coefficient of the camera based on the target detection requirements and the preliminarily estimated perspective coverage; A scene parameter initialization module is used to initialize the camera's view angle control parameters based on the camera's view angle adjustment coefficient and determine the current scene characteristics based on preset scene sensor recognition; An effect evaluation module, configured to evaluate the camera's viewing angle coverage effect based on the current scene characteristics and the viewing angle control parameters of the initialized camera; A model optimization module is used to establish a dynamic view angle control optimization model based on deep learning, analyze the current scene characteristics and the change trend of the view angle coverage effect of the camera by the view angle control parameters of the initialized camera, and generate the optimal view angle control parameters; The optimization control module is used to optimize and control the viewing angle operating parameters of the camera based on the optimal viewing angle control parameters.

[0016] In addition, to achieve the above-mentioned purpose, the present invention also provides a network camera, which includes: a memory, a processor, and a deep learning-based camera dynamic viewing angle adjustment program stored on the memory and runnable on the processor, and the deep learning-based camera dynamic viewing angle adjustment program is configured to implement the steps of the deep learning-based camera dynamic viewing angle adjustment method as described in any one of the above texts.

[0017] In addition, to achieve the above-mentioned purpose, the present invention also provides a storage medium, on which a deep learning-based camera dynamic viewing angle adjustment program is stored. When the deep learning-based camera dynamic viewing angle adjustment program is executed by a processor, the steps of the deep learning-based camera dynamic viewing angle adjustment method as described in any one of the above texts are implemented.

[0018] The present invention provides a method for dynamic camera viewing angle adjustment based on deep learning. This method breaks the limitations of traditional fixed viewing angles by integrating scene types, target detection requirements, and real-time scene features, and realizes dynamic adaptive adjustment of camera viewing angles. It effectively copes with complex scenes such as indoor and outdoor switching, dynamic tracking, and multi-target detection, and avoids blind spots or redundant coverage. Based on the quantitative correlation between target detection requirements and viewing angle coverage, the adjustment coefficient and deep learning model are iteratively optimized to accurately balance the viewing angle range and detection accuracy, avoid invalid coverage or detection omissions caused by fixed parameters, and improve the efficiency of monitoring resource utilization. By using deep learning to model the nonlinear relationship between scene features and viewing angle parameters, the changing trend of viewing angle coverage effect is dynamically analyzed to generate optimal control parameters, solve the complex scene optimization problem that is difficult to handle with traditional regularized algorithms, and significantly improve the integrity and recognition accuracy of target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the network camera structure of the hardware operating environment involved in the embodiment of the present invention; Figure 2 This is a flow chart of an embodiment of a method for dynamic camera viewing angle adjustment based on deep learning according to the present invention; Figure 3 This is a structural block diagram of an embodiment of a camera dynamic viewing angle adjustment device based on deep learning of the present invention.

[0020] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0021] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0022] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a network camera in the hardware operating environment involved in an embodiment of the present invention.

[0023] like Figure 1 As shown, the network camera may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen. Optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. In the present invention, the wired interface of the user interface 1003 may be a USB interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a wireless fidelity (WI-FI) interface). The memory 1005 may be a high-speed random access memory (RAM) or a non-volatile memory (NVM), such as a disk drive. The memory 1005 may also be a storage device independent of the processor 1001.

[0024] Those skilled in the art will understand that Figure 1 The structure shown in the figure does not constitute a limitation to the network camera, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0025] like Figure 1As shown, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a camera dynamic viewing angle adjustment program based on deep learning.

[0026] exist Figure 1 In the network camera shown, the network interface 1004 is mainly used to connect to the background server and communicate data with the background server; the user interface 1003 is mainly used to connect to peripheral devices; the network camera calls the deep learning-based camera dynamic viewing angle adjustment program stored in the memory 1005 through the processor 1001, and executes the deep learning-based camera dynamic viewing angle adjustment method provided by the embodiment of the present invention.

[0027] Based on the above hardware structure, an embodiment of the camera dynamic viewing angle adjustment method based on deep learning of the present invention is proposed.

[0028] Reference Figure 2 , Figure 2 This is a flow chart of an embodiment of a method for dynamic viewing angle adjustment of a camera based on deep learning of the present invention, and proposes an embodiment of a method for dynamic viewing angle adjustment of a camera based on deep learning of the present invention.

[0029] In one embodiment, the method for dynamic camera viewing angle adjustment based on deep learning includes the following steps: S10: Obtain the current working scene type and target detection requirements of the camera.

[0030] S20: Preliminarily estimating the camera's viewing angle coverage according to the current working scene type, and determining the camera's viewing angle adjustment coefficient based on the target detection requirements and the preliminarily estimated viewing angle coverage.

[0031] It should be noted that the current working scene type refers to the functional attributes and environmental characteristics of the camera's current environment, including but not limited to the scene's purpose, spatial attributes, dynamic characteristics, and lighting conditions. Scene purposes include security monitoring, industrial inspection, and human-computer interaction; spatial attributes can include indoor / outdoor, open area / narrow area; dynamic characteristics can include static scenes / dynamic tracking scenes; and lighting conditions can include bright light / low light / nighttime. Object detection requirements are the specific requirements that the user or system has for the camera to detect targets, including target type, detection accuracy, detection range, and real-time detection. Target types can include people, vehicles, and equipment parts; detection accuracy can include requirements for minimum pixel size; detection range can include global coverage / focused areas; and real-time detection requirements include high-frequency detection of high-speed moving targets. Field of view coverage refers to the physical spatial range that can be clearly imaged by the camera's current field of view. This is determined by lens parameters, mounting position, and imaging resolution, and can be quantified using horizontal / vertical coverage distances, effective detection radius, or coverage area. Lens parameters include focal length and field of view, while mounting position includes height and tilt angle. The perspective adjustment coefficient quantifies the correction parameters required for target detection to the initial perspective. By integrating scene type constraints with detection priority, it generates adjustment weights for parameters such as focal length and rotation angle. The perspective adjustment coefficient can include a range correction coefficient, a precision correction coefficient, and a dynamic response coefficient. The range correction coefficient is used to expand or narrow the perspective range; the precision correction coefficient is used to improve local detection resolution; and the dynamic response coefficient is used to adapt to the target's motion speed.

[0032] In practice, the current working scene type can be obtained by acquiring basic data such as ambient light intensity and camera installation inclination through the camera's built-in or external sensors. External sensors include light sensors and IMU inertial sensors. Pre-trained deep learning models are used to classify real-time images and identify scene types, such as indoor warehouses or intersections, while also detecting dynamic features of the scene, such as pedestrian density and vehicle flow. Target detection requirements can be obtained by setting the target type, minimum detection size, and detection priority through an interactive interface. The target type can be pedestrian detection; the minimum detection size can be target pixels ≥50×50; the detection priority can be to prioritize the lower right corner of the image. Alternatively, requirements can be obtained from upper-level algorithm modules through system task dispatching. For example, a target tracking system requires the target to be kept in the center of the image and occupy ≥10%.

[0033] It should be understood that the preliminary estimation of the viewing angle coverage based on the scene type can be matched through a preset scene library, for example, a mapping table of typical scenes and initial viewing angle parameters is established. For example, in an indoor conference room scene, the preset focal length is 8mm, the horizontal field of view angle is 40°, and the coverage distance is 5-10m; or combined with real-time environmental parameter correction, if the installation height is detected to be 6m, which is higher than the preset value of 3m, the coverage distance estimate is expanded. In this embodiment, the viewing angle adjustment coefficient includes a range correction coefficient and an accuracy correction coefficient β; if the target detection requirement is global coverage, α≥1, and the viewing angle is expanded; if the requirement is high-precision detection of key areas, α<1, and α can be the ratio of the required coverage area of ​​the target to the preliminary estimated coverage area. The accuracy correction coefficient β can calculate the required camera focal length based on the minimum pixel size of the target.

[0034] S30: Initializing the camera's viewing angle control parameters based on the camera's viewing angle adjustment coefficient, and determining current scene features based on preset scene sensor recognition.

[0035] It should be noted that a camera's view control parameters are adjustable parameters used to directly control the camera's physical view angle or imaging range. They are the core execution variables for view adjustment. Initializing these parameters preliminarily sets the camera's operating state. A camera's view control parameters can include both physical adjustment parameters and imaging control parameters. Physical adjustment parameters include focal length, horizontal / vertical rotation angles, and tilt angles. Imaging control parameters include the digital zoom factor and ROI (Region of Interest). The digital zoom ratio (Digital Zoom Ratio) is a non-physical adjustment that achieves localized magnification by cropping the image. The ROI (Region of Interest) is the coordinate of the region of interest (ROI), which defines the camera's preferred detection area, such as the center of the image. As the execution interface for view adjustment, the camera's view control parameters interact directly with the hardware, mapping adjustment coefficients to specific parameter values, such as increasing the focal length or rotation angle based on the range correction factor.

[0036] Pre-installed scene sensors are hardware sensors pre-deployed on the camera device or its surroundings to collect real-time scene characteristics, providing the system with multimodal input information in addition to visual data. Pre-installed scene sensors include environmental state sensors, motion / presence sensors, and device state sensors. Environmental state sensors include light sensors, temperature sensors, and humidity / pressure sensors. Motion / presence sensors include PIR (passive infrared) sensors and ultrasonic / millimeter-wave radars. PIR (passive infrared) sensors detect the presence of moving objects in an area and determine scene dynamics. Ultrasonic / millimeter-wave radars are used for high-precision detection of target distance and speed, such as high-speed robotic arms in industrial scenarios. Device state sensors include IMUs (inertial measurement units) and lens parameter sensors. IMUs integrate accelerometers and gyroscopes to obtain the camera's current posture parameters. Lens parameter sensors provide real-time feedback on the actual values ​​of physical parameters such as focal length and rotation angle, which are used for parameter calibration in closed-loop control.

[0037] Specifically, before initializing the camera's viewing angle control parameters based on the camera's viewing angle adjustment coefficient and determining the current scene features based on preset scene sensor recognition, the method further includes: Obtaining a working scene type of the camera and several scene features of the working scene type, and determining a viewing angle control range corresponding to the scene features of the working scene type, wherein the scene features of the working scene type include scene complexity, target motion speed, and target size; Accordingly, the camera viewing angle adjustment coefficient is based on the camera, the viewing angle control parameter of the camera is initialized, and the current scene feature is determined based on the preset scene sensor recognition, including: The viewing angle adjustment algorithm is used to correct the viewing angle control range corresponding to the scene characteristics of the working scene type based on the viewing angle adjustment coefficient of the camera, determine the viewing angle control parameters of the initialized camera, and determine the current scene characteristics based on the preset scene sensor recognition.

[0038] It should be noted that the scene characteristics of a work scene type refer to key quantitative indicators that describe the core attributes of the camera's work scene. They are used to characterize the constraints and requirements of the scene on view adjustment. They include three core characteristics: scene complexity, target motion speed, and target size. Scene complexity is a comprehensive indicator that measures the number of targets in the scene, the complexity of the spatial layout, and interference factors, and is used to characterize the difficulty of view adjustment. Target motion speed is the average speed of targets in the scene and is used to determine whether rapid view adjustment is required to maintain target tracking. For example, high-speed moving targets require a faster response rate for view control parameters, such as reducing the focal length to expand the field of view to prevent the target from moving out of the frame. Target size is the actual size of the target in physical space or the pixel size in the image, such as the proportion of pixels in the frame occupied by the target. The view control range refers to the pre-set adjustable range of the camera view control parameters for a specific work scene type and its scene characteristics. It is used to limit the reasonable range of initialization parameters to avoid ineffective or dangerous view adjustments. The perspective adjustment algorithm is a mathematical model or rule engine used to associate the perspective adjustment coefficient with the perspective control range. By integrating scene feature constraints and detection requirements, it dynamically modifies the initial perspective control range and generates executable initialization parameters. Specifically, the preset control range is expanded or reduced based on the adjustment coefficient, such as modifying the focal length range from 8mm-25mm to 6mm-20mm to accommodate a wider field of view. When the target size requires a long focal length and the target movement speed requires a short focal length, a compromise range is generated through weight distribution.

[0039] It should be understood that traditional solutions rely solely on scene type labels, such as indoor / outdoor. This embodiment deconstructs scene requirements into quantifiable parameter constraints based on three core characteristics: scene complexity, target speed, and target size. For example, target size = 0.5m × 0.3m, target speed = 2m / s. This shifts perspective adjustment from empirical matching to data-driven, precise calculation. For example, in a parking lot scenario, if the detected target size is a large truck (5m × 2m) with a low speed (≤1m / s), a wide-angle perspective is prioritized for full coverage. If the detected target size is a small car (4m × 1.8m) with a high speed (≥5m / s), the field of view is narrowed and the focal length is increased to ensure continuous, clear imaging of high-speed targets. Furthermore, by presetting the perspective control range corresponding to scene characteristics, such as focal length 8-25mm and horizontal rotation ±90°, initialization parameters are ensured to meet detection requirements while remaining within the camera hardware capabilities. This prevents focus adjustment beyond the physical range of the lens and mechanical damage to the gimbal due to rotation angles.

[0040] The expression of the viewing angle adjustment algorithm is:

[0041] Where, To initialize the perspective control parameters of the camera, is the maximum perspective control range corresponding to the scene features of the working scene type, is the minimum perspective control range corresponding to the scene features of the working scene type, and E is the perspective adjustment coefficient of the camera.

[0042] It should be noted that the formula is within the preset parameter range , , and linear interpolation is performed through the adjustment coefficient E so that the final parameter falls within a reasonable range and at the same time meets the adjustment direction after demand quantification. When E = 0, = , indicating that the minimum control parameter corresponding to the scene features is adopted, such as the minimum focal length, corresponding to the maximum field of view angle, for maximizing the coverage range. When E = 1, = , indicating that the maximum control parameter corresponding to the scene features is adopted, such as the maximum focal length, corresponding to the minimum field of view angle, for maximizing the detail clarity. When 0 < E < 1, the parameter falls at the midpoint of the interval. For example, when E = 0.5, is the midpoint value of the range, suitable for scenes that balance the coverage range and details.

[0043] It should be understood that assuming the focal length control range of a certain scene is = 8mm (wide-angle end), = 25mm (telephoto end). If E = 0.8 and the demand is to magnify the details, = 8 + 0.8×(25 - 8) = 21.6mm, the focal length shifts towards the telephoto end, reducing the field of view angle to magnify the target imaging. If E = 0.3 and the demand is to expand the coverage, V = 8 + 0.3×(25 - 8) = 13.1mm, the focal length shifts towards the wide-angle end, expanding the field of view angle to cover more areas. In this embodiment, no matter how E changes, V is always restricted within , , avoiding exceeding the hardware capabilities, such as the focal length adjustment range and the pan / tilt rotation limit, ensuring the parameter feasibility from a mathematical perspective. Using a linear model instead of a complex non-linear function is convenient for engineering implementation and parameter debugging, and the physical meaning of the adjustment coefficient E is intuitive. For example, for every increase of 0.1, the focal length increases by 1.7mm, reducing the system deployment difficulty. It is applicable to the adjustment of a single perspective control parameter, such as the focal length and the horizontal rotation angle, or as a basic unit for multi-parameter adjustment. The formula is independently applied to each parameter, and then fused through weights.

[0044] S40: Evaluate the perspective coverage effect of the camera according to the current scene features and the perspective control parameters of the initialized camera.

[0045] It's important to note that view coverage refers to the degree to which the camera's current view control parameters adapt to the target scene. It is a key indicator of a camera's ability to effectively complete monitoring / detection tasks. This performance is quantitatively evaluated by comparing the required scene features with the actual imaging results, and directly impacts target detection, tracking, and recognition performance.

[0046] Specifically, the evaluating the camera's viewing angle coverage effect according to the current scene features and the initialization camera's viewing angle control parameters includes: Assigning a coverage risk score to each scene feature according to a coverage risk change trend that occurs when several scene features in the current scene feature affect camera control, thereby obtaining a scene feature coverage risk score; Standardizing the scenario feature coverage risk score through a standardized formula; Based on the standardized scene feature coverage risk scores, determine the relative importance of each scene feature coverage risk score to the view coverage during camera control, assign a coverage risk weight to each scene feature, and obtain the scene feature coverage risk weight; Calculating a camera's coverage weighted score based on the scene feature coverage risk score and the scene feature coverage risk weight; Based on the coverage weighted score of the camera and the initialization camera's viewing angle control parameters, the camera's viewing angle coverage effect is calculated through a coverage evaluation formula.

[0047] It should be noted that the coverage risk trend refers to the evolution of the coverage risk of camera view control when a scene feature changes. For example, an increase in target speed may increase the risk of losing the target, indicating a positive risk trend; a decrease in scene complexity may decrease the risk of background interference, indicating a negative risk trend. By mapping scene features to coverage risk, we can identify the potential threats to view control posed by feature changes and provide a logical basis for risk scoring. The coverage risk score is a quantitative score of the degree of view coverage risk posed by each scene feature, ranging from [0, 100] to [0, 1]. A higher score indicates a greater coverage risk posed by that feature. Feature values ​​can be mapped to a risk score based on preset rules or machine learning models. Quantifying the threat level of a single feature to view control serves as the basis for subsequent weighting and comprehensive assessment. The scene feature coverage risk score is a coverage risk score calculated for each specific scene feature, forming a scoring vector [S1, S2, S3], where Si corresponds to the risk score for the i-th scene feature. Multidimensional scene features are converted into computable risk quantification indicators. For example, the scoring vector for a parking lot scene is [70, 50, 80], corresponding to the risk scores for target speed, scene complexity, and target size, respectively. A standardized formula is used to convert risk scores for scene features with different dimensions and value ranges into a unified mathematical method, eliminating the interference of dimensional differences in weight calculation. This ensures that scores for different features are comparable and prevents features with large value ranges from dominating weight calculations.

[0048] The coverage risk weight reflects the relative importance of each scene feature to the camera's field of view coverage. Its value range is [0, 1], and the sum of all feature weights is 1. A higher weight indicates a more critical impact of the feature on field of view control. It also reflects the priority of scene characteristics over field of view control. For example, in live sports events, target speed is weighted higher than scene complexity, prioritizing tracking of high-speed athletes. In static object detection, target size is weighted higher, prioritizing detail clarity. The scene feature coverage risk weight assigns a coverage risk weight to each specific scene feature, forming a weight vector [W1, W2, W3] that corresponds one-to-one with the score vector. The weighted coverage score is a comprehensive risk score calculated by combining the scene feature coverage risk scores with the corresponding weights. Through weighted summation, this multi-dimensional risk is transformed into a single comprehensive indicator reflecting the overall threat level of all scene features to field of view coverage. For example, a higher weighted score indicates that the current field of view parameters are less likely to meet the scene requirements. The coverage evaluation formula combines the weighted coverage score with the initialized field of view control parameters to form a mathematical expression for calculating field of view coverage. This can be a function involving linear combinations, nonlinear mappings, or machine learning models.

[0049] Wherein, the coverage evaluation formula is:

[0050] Where C is the camera's viewing angle coverage effect, is the risk weight of the camera’s i-th scene feature coverage, is the coverage risk score of the i-th scene feature of the camera, V is the initialization camera's viewing angle control parameter, and n is the total number of scene features of the camera.

[0051] It should be noted that It is a weighted sum model that aggregates the risk scores of multiple scenario features into a comprehensive risk index according to their importance. The risk score of each feature is Reflect its own risk, such as the score of the target being too small to be identified, the weight This reflects the priority of the risk in the overall assessment. For example, in security scenarios, the weight of target size is higher than that of scene brightness. The higher the weighted sum, the greater the threat posed by the overall scene features to the viewing angle coverage. For example, a combination of a complex scene and a high-speed target will result in a high overall risk. In this embodiment, the initialization of the camera's viewing angle control parameter V is used to associate the initialization viewing angle parameter with the scene risk. If V is the focal length, that is, the larger the value, the smaller the field of view but the clearer the details. When the scene feature is that the target is too small, that is, the risk score is high, a larger V can reduce the risk and magnify the target imaging. At this time, the product of V and the risk-weighted sum represents the actual response effect of the current parameter to the risk.

[0052] S50: Establish a dynamic perspective control optimization model based on deep learning, analyze the current scene features and the perspective control parameters of the initialized camera on the camera's perspective coverage effect change trend, and generate optimal perspective control parameters.

[0053] S60: Optimizing and controlling the viewing angle operating parameters of the camera based on the optimal viewing angle control parameters.

[0054] It should be noted that the dynamic view control optimization model can be an intelligent model that integrates deep learning technology. By learning the mapping relationship between historical scene characteristics and view control parameters, it dynamically predicts and optimizes camera view parameters to maximize view coverage. Its core is to leverage the nonlinear fitting capabilities of neural networks, replacing traditional rule-based engines to achieve end-to-end automated optimization. The view coverage trend is the evolution of view coverage over time or with parameter adjustments as scene characteristics or view control parameters change.

[0055] Optimal view control parameters are the combination of view control parameters that achieve the global or local optimal camera view coverage under the current scene characteristics. Dynamic view control adjusts the camera view parameters in real time based on changing scene characteristics, ensuring optimal coverage of key areas / targets. Optimal control is the process of converting the optimal view parameters generated by the model into actual camera execution instructions, including parameter mapping, execution verification, and feedback loops.

[0056] Specifically, the establishment of a dynamic view control optimization model based on deep learning, analyzing the change trend of the view coverage effect of the camera by the current scene features and the view control parameters of the initialized camera, and generating the optimal view control parameters, includes: The viewing angle control range corresponding to the scene characteristics of the work scene type is used as an adjustable viewing angle control restriction condition; Construct a dynamic view control optimization model based on deep learning; Based on the deep learning-based dynamic perspective control optimization model, under the adjustable perspective control restriction condition, the current scene features and the perspective control parameters of the initialized camera are used as independent variable inputs to determine the changing trend of the perspective coverage effect of the camera. The perspective control parameters are gradually increased according to the adjustable perspective control restriction condition, and several rounds of iterations are performed to output the perspective control parameters that maximize the perspective coverage effect of the camera as the dependent variable, thereby generating the optimal perspective control parameters.

[0057] It should be noted that the adjustable view control constraints are pre-set feasible ranges for camera view control parameters based on the work scenario type and hardware physical characteristics. They are used to constrain the view parameters generated by the model to a safe and effective range. Independent variable inputs are variables that serve as model inputs in the deep learning model and are not affected by the model output. Here, they are scene features related to the camera view coverage effect, the initial view parameters, and information on their changing trends in coverage effect. Dependent variable outputs are target variables in the deep learning model that depend on the independent variable inputs and are optimized by the model. Here, they refer to the optimal view control parameters that maximize the camera view coverage effect while satisfying the adjustable view control constraints. Stepwise incrementation is a strategy that, during the iterative optimization process of the model, fine-tunes the view control parameters or constraints in stages according to a preset parameter adjustment step size or constraint relaxation strategy, gradually approaching the optimal solution. Iterations are the process of gradually optimizing the view control parameters through multiple rounds of model inference and parameter adjustment.

[0058] The expression of the dynamic viewing angle control optimization model is:

[0059] Where, is the optimal viewing angle control parameter, For the The viewing angle control parameters during round iteration, is the learning efficiency step size, is the gradient of the viewing angle control parameter, is the camera's view coverage effect function, V is the initialization view control parameter, S is the current scene feature, is the upper limit of the viewing angle control range corresponding to the scene characteristics of the work scene type, is the lower limit.

[0060] It should be noted that It is the perspective control parameter of the current iteration, that is, the independent variable in the formula. The upper limit of the viewing angle control range is the maximum parameter value vector determined by the working scene type and hardware characteristics. is the lower limit of the viewing angle control range, and similarly is the parameter minimum value vector. Ensure that the view control parameter V is always within a safe and feasible range to avoid damage to camera hardware or coverage failure due to parameter out-of-bounds. The inequality is essentially an explicit constraint on the parameter space and a mathematical expression of the adjustable view control restrictions in engineering implementation. is the optimal viewing angle control parameter, that is, the target parameter generated through iterative optimization. For the The viewing angle control parameters during the round iteration, where is the current round, j=1,2,…,m represents the historical parameters of m rounds back, reflecting the aggregation of historical iteration information. It is the historical parameter summation term, which represents the accumulation of parameters in the past m rounds. To improve learning efficiency, control the amplitude of each parameter update to avoid excessive adjustment amplitude leading to oscillation or too small adjustment amplitude leading to slow convergence. The gradient of the view coverage effect function with respect to the view parameter, i.e., the partial derivative vector of C with respect to V, indicates the direction and rate at which parameter adjustment affects the coverage effect. C(V, S) is the view coverage effect function, which takes the view parameter V and scene feature S as input and outputs the coverage effect evaluation value.

[0061] In addition, an embodiment of the present invention also proposes a storage medium, on which a deep learning-based camera dynamic viewing angle adjustment program is stored. When the deep learning-based camera dynamic viewing angle adjustment program is executed by a processor, the steps of the deep learning-based camera dynamic viewing angle adjustment method described above are implemented.

[0062] In addition, refer to Figure 3 The embodiment of the present invention further provides a device for adjusting the dynamic viewing angle of a camera based on deep learning, and the device for adjusting the dynamic viewing angle of a camera based on deep learning includes: The scene acquisition module 10 is used to obtain the current working scene type and target detection requirements of the camera; A perspective preliminary estimation module 20 is configured to preliminarily estimate the perspective coverage of the camera according to the current working scene type, and determine a perspective adjustment coefficient of the camera based on the target detection requirements and the preliminarily estimated perspective coverage; The scene parameter initialization module 30 is used to initialize the camera's view angle control parameters based on the camera's view angle adjustment coefficient and determine the current scene characteristics based on the preset scene sensor recognition; An effect evaluation module 40 is used to evaluate the camera's viewing angle coverage effect based on the current scene characteristics and the viewing angle control parameters of the initialized camera; A model optimization module 50 is configured to establish a dynamic view angle control optimization model based on deep learning, analyze the current scene characteristics and the change trend of the camera's view angle coverage effect of the view angle control parameters of the initialized camera, and generate optimal view angle control parameters; The optimization control module 60 is used to optimize and control the viewing angle operating parameters of the camera based on the optimal viewing angle control parameters.

[0063] Other embodiments or specific implementation methods of the deep learning-based camera dynamic viewing angle adjustment device described in the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0064] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0065] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent superiority or inferiority of the embodiments. In a unit claim that enumerates several means, several of these means may be embodied by the same item of hardware. The use of the terms first, second, and third, etc., does not denote any order; these terms should be interpreted as designations.

[0066] Through the above description of the embodiments, those skilled in the art will clearly understand that the above-mentioned embodiments and methods can be implemented using software plus the necessary general hardware platform. Of course, hardware can also be used, but in many cases, the former is a more preferred embodiment. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory image (ROM) / random access memory (RAM), a magnetic disk, or an optical disk) and includes a number of instructions for enabling an end-user device (such as a mobile phone, computer, server, air conditioner, or network user device, etc.) to execute the methods described in various embodiments of the present invention.

[0067] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for dynamic camera viewing angle adjustment based on deep learning, characterized in that: The method comprises: Get the camera's current working scene type and target detection requirements; Preliminarily estimating the camera's viewing angle coverage according to the current working scene type, and determining the camera's viewing angle adjustment coefficient based on the target detection requirements and the preliminarily estimated viewing angle coverage; Initializing the camera's viewing angle control parameters based on the camera's viewing angle adjustment coefficient, and determining current scene features based on preset scene sensor recognition; Evaluating the camera's viewing angle coverage effect based on the current scene features and the viewing angle control parameters of the initialized camera; Establishing a dynamic view control optimization model based on deep learning, analyzing the change trend of the view coverage effect of the camera based on the current scene characteristics and the view control parameters of the initialized camera, and generating the optimal view control parameters; The viewing angle operating parameters of the camera are optimized and controlled based on the optimal viewing angle control parameters.

2. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 1, wherein: Before initializing the camera's viewing angle control parameters based on the camera's viewing angle adjustment coefficient and determining the current scene features based on preset scene sensor recognition, the method further includes: Obtaining a working scene type of the camera and several scene features of the working scene type, and determining a viewing angle control range corresponding to the scene features of the working scene type, wherein the scene features of the working scene type include scene complexity, target motion speed, and target size; Accordingly, the camera viewing angle adjustment coefficient is based on the camera, the viewing angle control parameter of the camera is initialized, and the current scene feature is determined based on the preset scene sensor recognition, including: The viewing angle adjustment algorithm is used to correct the viewing angle control range corresponding to the scene characteristics of the working scene type based on the viewing angle adjustment coefficient of the camera, determine the viewing angle control parameters of the initialized camera, and determine the current scene characteristics based on the preset scene sensor recognition.

3. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 2, wherein: The expression of the viewing angle adjustment algorithm is: Where, To initialize the camera's viewing angle control parameters, is the maximum viewing angle control range corresponding to the scene characteristics of the work scene type, is the minimum viewing angle control range corresponding to the scene characteristics of the working scene type, and E is the viewing angle adjustment coefficient of the camera.

4. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 3, wherein: The evaluating the camera's viewing angle coverage effect according to the current scene feature and the initialization camera's viewing angle control parameter includes: Assigning a coverage risk score to each scene feature according to a coverage risk change trend that occurs when several scene features in the current scene feature affect camera control, thereby obtaining a scene feature coverage risk score; Standardizing the scenario feature coverage risk score through a standardized formula; Based on the standardized scene feature coverage risk scores, determine the relative importance of each scene feature coverage risk score to the view coverage during camera control, assign a coverage risk weight to each scene feature, and obtain the scene feature coverage risk weight; Calculating a camera's coverage weighted score based on the scene feature coverage risk score and the scene feature coverage risk weight; Based on the coverage weighted score of the camera and the initialization camera's viewing angle control parameters, the camera's viewing angle coverage effect is calculated through a coverage evaluation formula.

5. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 4, wherein: The coverage evaluation formula is: Where C is the camera's viewing angle coverage effect, is the risk weight of the camera’s i-th scene feature coverage, is the coverage risk score of the i-th scene feature of the camera, V is the initialization camera's viewing angle control parameter, and n is the total number of scene features of the camera.

6. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 5, wherein: The method of establishing a dynamic view angle control optimization model based on deep learning, analyzing the change trend of the view angle coverage effect of the camera based on the current scene characteristics and the view angle control parameters of the initialized camera, and generating the optimal view angle control parameters includes: The viewing angle control range corresponding to the scene characteristics of the work scene type is used as an adjustable viewing angle control restriction condition; Construct a dynamic view control optimization model based on deep learning; Based on the deep learning-based dynamic perspective control optimization model, under the adjustable perspective control restriction condition, the current scene features and the perspective control parameters of the initialized camera are used as independent variable inputs to determine the changing trend of the perspective coverage effect of the camera. The perspective control parameters are gradually increased according to the adjustable perspective control restriction condition, and several rounds of iterations are performed to output the perspective control parameters that maximize the perspective coverage effect of the camera as the dependent variable, thereby generating the optimal perspective control parameters.

7. The method for dynamic camera viewing angle adjustment based on deep learning according to claim 6, wherein: The expression of the dynamic viewing angle control optimization model is: Where, is the optimal viewing angle control parameter, For the The viewing angle control parameters during round iteration, is the learning efficiency step size, is the gradient of the viewing angle control parameter, is the camera's view coverage effect function, V is the initialization view control parameter, S is the current scene feature, is the upper limit of the viewing angle control range corresponding to the scene characteristics of the work scene type, is the lower limit.

8. A camera dynamic viewing angle adjustment device based on deep learning, characterized in that: The camera dynamic viewing angle adjustment device based on deep learning includes: The scene acquisition module is used to obtain the camera's current working scene type and target detection requirements; A perspective preliminary estimation module, configured to preliminarily estimate the perspective coverage of the camera according to the current working scene type, and determine the perspective adjustment coefficient of the camera based on the target detection requirements and the preliminarily estimated perspective coverage; A scene parameter initialization module is used to initialize the camera's view angle control parameters based on the camera's view angle adjustment coefficient and determine the current scene characteristics based on preset scene sensor recognition; An effect evaluation module, configured to evaluate the camera's viewing angle coverage effect based on the current scene characteristics and the viewing angle control parameters of the initialized camera; A model optimization module is used to establish a dynamic view control optimization model based on deep learning, analyze the current scene characteristics and the change trend of the view coverage effect of the camera's view control parameters of the initialized camera, and generate the optimal view control parameters; The optimization control module is used to optimize and control the viewing angle operating parameters of the camera based on the optimal viewing angle control parameters.

9. A network camera, characterized in that: The network camera includes: a memory, a processor, and a deep learning-based camera dynamic viewing angle adjustment program stored in the memory and runnable on the processor, wherein the deep learning-based camera dynamic viewing angle adjustment program is configured to implement the steps of the deep learning-based camera dynamic viewing angle adjustment method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a deep learning-based camera dynamic viewing angle adjustment program, and when the deep learning-based camera dynamic viewing angle adjustment program is executed by the processor, the steps of the deep learning-based camera dynamic viewing angle adjustment method as described in any one of claims 1 to 7 are implemented.

Images