Abnormal condition detection method and system for intelligent safety help-seeking cloud platform

Abstract

The invention provides an abnormal condition detection method and system for an intelligent safety help-seeking cloud platform, and the method comprises the steps: firstly determining a distance condition between a standard calibration object and a camera device in a visual field, and determining the size of a prior frame corresponding to the distance according to the average size of a key detection object; and adjusting the initialization grid based on the determined size of the prior frame, using the adjusted prior frame as an initialization parameter of training of a target identification model for model training, and finally carrying out abnormal condition detection on a real-time image acquired by a camera device based on the trained model. The invention further provides an adaptive sample expansion scheme. According to the scheme, the characteristics of the image collected under the hardware architecture of the intelligent safety help-seeking cloud platform are utilized, the prior frame of the target detection algorithm is optimized, the algorithm complexity is reduced, and the training efficiency and prediction accuracy of the target detection model are improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and more specifically, to an abnormal condition detection method and system for a smart security help cloud platform. Background Art

[0002] With the rapid development of artificial intelligence and automation technology, the smart security assistance cloud platform is playing an increasingly extensive and important role in security monitoring, danger alarm, intelligent treatment, and smart city management.

[0003] The basic implementation principle of the common security assistance platform on the market currently adopts a computing center to multiple monitoring terminal architecture, and the background obtains the real-time data of the terminal for data processing. If the target data is recognized, an alarm or rescue mechanism is triggered. This system has a wide range of application scenarios, such as factory management, warehouse management, traffic control, video surveillance and other special customizations. Its core technology is mainly content recognition algorithm, including personnel information acquisition / target recognition through camera devices, face recognition, target behavior and action recognition, scene recognition, dangerous situation recognition, etc. At present, the core technology route of content recognition technology is tending towards machine self-learning image processing based on neural network algorithms. For example, the mainstream target recognition algorithms are divided into TwoStage type algorithms and OneStage type algorithms. The former classic algorithms include: R-CNN / FR-CNN, etc., and the latter classic algorithms include YOLO series algorithms / SSD algorithms, etc.

[0004] However, the above algorithm still has disadvantages when applied to the smart security help cloud platform. First, the image data of the smart security help cloud platform has its own particularity compared with conventional image data. For example, (1) the recognition size of distant objects is generally smaller than that of nearby objects; (2) the types of monitored target objects are relatively fixed; (3) the shooting has relatively fixed angle characteristics; (4) and other special characteristics of special scenes; therefore, if the algorithm is not improved based on the data particularity when processing image data, the model calculation efficiency is not high and the data processing accuracy is not high; secondly, the existing image training data for target recognition is not compatible with the characteristics of this field, resulting in insufficient image training data in this field, insufficient model training in practical applications, and low data processing accuracy; thirdly, the model training algorithm is complex, and most terminal devices do not have a training environment for complex algorithms, which is insufficient to provide hardware computing power support for complex algorithms, resulting in model simplification and weakening. In order to be applicable to various types of unknown target objects, the target recognition algorithm in the prior art adopts a broad data training method, which is an important reason for the low efficiency of its algorithm. Therefore, it is necessary to improve the target recognition algorithm based on its data particularity in the field of smart security help technology.

[0005] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as acknowledging or suggesting in any form that the information constitutes the prior art already known to those skilled in the art. Summary of the invention

[0006] In view of the problems existing in the prior art, the present invention provides an abnormal condition detection method and system for a smart security help cloud platform.

[0007] Specifically, the present application claims an abnormal condition detection method for a smart security help cloud platform, the method comprising the steps of:

[0008] S100, determining a standard calibration object, moving the standard calibration object so that it traverses the field of view of the camera device, and in the process of traversing the field of view, executing the step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device;

[0009] S110, setting a priori box sizes corresponding to a plurality of distances according to an average size parameter of an object to be detected by the camera device;

[0010] S120, performing initialization grid division on the image acquired by the camera device, determining distance data from the midpoint of the bottom edge of the grid to the camera device, adjusting the size parameter of the initialization grid according to the size of the priori frame corresponding to the distance data to obtain an adjusted priori frame, and using the adjusted priori frame as an initialization parameter for training the target recognition model for model training;

[0011] S130: The computing device performs abnormal condition detection on the real-time image acquired by the camera device based on the trained target recognition model.

[0012] Furthermore, the standard calibration object is moved to traverse the field of view of the camera device. During the traversal of the field of view, the distance calculation is performed for each pixel point to calculate the actual distance D between the currently traversed pixel point and the camera device:

[0013]

[0014] In the above formula (1), L is the size of the standard calibration object, f is the focal length of the camera device, and d is the image size of the standard calibration object in the camera device.

[0015] Furthermore, a scene includes multiple types of key detection objects, and the prior box size corresponding to the distance data includes multiple mappings, each mapping is associated with each key detection object.

[0016] Furthermore, it is applied to real-time monitoring scenarios, where the key detection objects are pedestrians, and the shape of the prior frame is the human figure ratio. When abnormal behavior detection is performed, the regular shape of the behavior in the image is determined based on the specific behavior content, and the initial shape of the prior frame is determined based on the regular shape.

[0017] Furthermore, an expansion operation is performed based on the target object calibrated in the existing training sample image. The expansion operation is specifically: cutting the target object based on the calibration frame, displacing it in the training sample image, determining the distance data from the pixel at this position to the camera device based on the position of the center of the bottom edge of the calibration frame at the end point of the displacement, determining the prior frame size of the position according to the mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training sample image based on the determined prior frame size.

[0018] Also provided is an abnormal condition detection system for a smart safety help cloud platform, comprising:

[0019] A distance determination module is used to determine a standard calibration object, move the standard calibration object so that it traverses the field of view of the camera device, and in the process of traversing the field of view, perform a step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device;

[0020] A priori frame mapping module, used to set the priori frame sizes corresponding to multiple distances according to the average size parameters of the key detection objects of the camera device;

[0021] A training module is used to initialize grid division of the image acquired by the camera device, determine the distance data from the midpoint of the bottom edge of the grid to the camera device, adjust the size parameters of the initialization grid according to the size of the prior frame corresponding to the distance data to obtain an adjusted prior frame, and use the adjusted prior frame as an initialization parameter for training the target recognition model for model training;

[0022] The detection module is a computing device that performs abnormal condition detection on the real-time image acquired by the camera device based on the trained target recognition model.

[0023] Furthermore, in the above system, the standard calibration object is moved to traverse the field of view of the camera device. During the traversal of the field of view, the distance calculation is performed for each pixel point to calculate the actual distance D between the currently traversed pixel point and the camera device:

[0024]

[0025] In the above formula (1), L is the size of the standard calibration object, f is the focal length of the camera device, and d is the image size of the standard calibration object in the camera device.

[0026] Furthermore, in the above system, a scene includes multiple types of key detection objects, and the prior box size corresponding to the distance data includes multiple mappings, each mapping is associated with each key detection object.

[0027] Furthermore, in the above system, it is applied to real-time monitoring scenarios, where the key detection objects are pedestrians, and the shape of the prior frame is the human figure ratio. When abnormal behavior detection is performed, the regular shape of the behavior in the image is determined based on the specific behavior content, and the initial shape of the prior frame is determined based on the regular shape.

[0028] Furthermore, in the above system, an expansion operation is performed based on the target object calibrated in the existing training sample image. The expansion operation is specifically: cutting the target object based on the calibration frame, displacing it in the training sample image, determining the distance data from the pixel at this position to the camera device based on the position of the center of the bottom edge of the calibration frame at the end point of the displacement, determining the prior frame size of the position according to the mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training sample image based on the determined prior frame size.

[0029] The present invention provides an abnormal condition detection method and system for a smart safety help cloud platform. The method first determines the distance between the standard calibration object and the camera device in the field of view, determines the size of the prior frame corresponding to the distance according to the average size of the key detection object, adjusts the initialization grid based on the determined prior frame size, uses the adjusted prior frame as the initialization parameter for training the target recognition model for model training, and finally detects abnormal conditions on the real-time image obtained by the camera device based on the trained model. The present invention also provides a corresponding sample expansion scheme. The above method of the present invention utilizes the image characteristics collected under the hardware architecture of the smart safety help cloud platform, optimizes the prior frame of the target detection algorithm, reduces the algorithm complexity, and improves the training efficiency and prediction accuracy of the target detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only used to illustrate some embodiments of the present application. For ordinary technicians in this field, without paying creative work, they can also obtain other technical features, connection relationships and even method steps not mentioned in the drawings based on these drawings.

[0031] Figure 1 It is a flow chart of an abnormal condition detection method for a smart security help cloud platform provided by an embodiment of the present invention;

[0032] Figure 2It is a structural diagram of an abnormal condition detection system for a smart security help cloud platform provided by an embodiment of the present invention; Specific embodiments

[0033] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0034] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein, for example. In addition, the terms "including" and "corresponding to" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0036] Embodiment 1: Figure 1 As shown, the present invention provides an abnormal condition detection method for a smart security help cloud platform, the method comprising the steps of:

[0037] S100, determining a standard calibration object, moving the standard calibration object so that it traverses the field of view of the camera device, and during the traversal process, executing the step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device;

[0038] For target detection tasks, classic algorithms usually include the concept of prior frames. The classic solution is to traverse all possible calibration frames on the input image, then select the correct target frame, and adjust (select) the position and size of the calibration frame based on conditional judgment and conditional iteration until it matches the object to be detected, thus completing the target detection task. The calibration frame for prediction will be adjusted later and is called a prior frame. The initialized prior frame has certain predictive properties. If its shape is similar to the target object, the number of adjustment iterations is small and the model recognition efficiency is high. On the contrary, if its shape is far from the target object, the number of adjustment iterations is large and the model recognition efficiency is low. In order to increase the execution efficiency of the target detection task, prior frames of different shapes / sizes are set at the same position, which will result in a large amount of model calculation.

[0039] The present application optimizes the setting method of the prior frame in many aspects. The first aspect is to determine the size of the prior frame. In order to determine the size of the prior frame, the present application performs distance analysis on the image captured by the camera device. First, determine a standard calibration object, which can be an object of known size, such as a staff member with a height of 1.75m / a shoulder width of 0.6m, or a work vehicle of known standard size. Move the standard calibration object to traverse the field of view of the camera device. In the process of traversing the field of view, perform distance calculation for each pixel point, that is, based on the size parameters of the standard calibration object, the imaging size of the standard calibration object in the camera device, and the physical parameters of the camera device, calculate the actual distance D of the currently traversed pixel point from the camera device:

[0040]

[0041] In the above formula (1), L is the size of the standard calibration object, 1 is the focal length of the camera device, and d is the imaging size of the standard calibration object in the camera device.

[0042] In specific implementations, the main methods for traversing the field of view of the camera device are: (1) downsampling traversal: perform N-interval sampling on the image pixels, determine the pixel position obtained by sampling, and move the standard calibration object to the pixel position to perform a traversal operation; (2) scene-based straight line sampling traversal: this method is usually applicable to linear fields of view similar to road scenes. The sampling line is set based on the direction of the road, the pixel position is determined based on the sampling line, and the standard calibration object is moved to the pixel position to perform a traversal operation; (3) scene-based endpoint sampling traversal: this method is usually applicable to irregular polygonal fields of view similar to squares. The sampling endpoints are set based on the shape of the irregular field of view, the endpoints are aligned to obtain sampling line segments, the pixel position is determined based on the sampling line segments, and the standard calibration object is moved to the pixel position to perform a traversal operation.

[0043] S11 O, setting the size of the priori frame corresponding to the plurality of distance data according to the average size parameter of the key detection object of the camera device;

[0044] Prior frame matching is a common step in the target detection algorithm. Its goal is to move and adjust the prior frame in the direction of the actual frame. The closer the prior frame is to the actual frame, the fewer iterations and the less calculations. Therefore, it is hoped that the prior frame is as close to the actual frame as possible. Since the size / shape of the focus object captured by the camera in different scenes is relatively uniform to a certain extent, based on this principle, the task of this step is to determine the size of the predicted prior frame on each pixel.

[0045] In one embodiment, the scene of the camera device is road traffic, and the key detection objects are different types of cars. The length of a family car is between 3800mm-4200mm and the width is about 1600mm; the length of a minibus is about 5990mm and the width is about 2050mm; the length of a bus is 8600mm and the width is 2300mm. In one embodiment, the application scene of the camera device is a park or square. At this time, the key detection object is pedestrians. The height of men is about 1.60-1.80 (m), and the height of women is about 1.50-1.70 (m). The overall characteristics of the height of the two are the same in graphic expression. In order to reduce the complexity of the algorithm, 1.7m high and 0.6m wide can be taken.

[0046] According to the average size parameter of the key detection object, set the prior frame sizes corresponding to different multiple distances. In the park or square scene, the key detection object is a pedestrian, whose average size parameter is 1.70m. Assuming that the image captured by the camera device is 720P, then after the traversal step of the standard calibration object in 1280*720 pixels, the distance between the determined pixel and the camera device in the scene is known, and the imaging size is further determined by formula (1). Then, based on the imaging size, the average size parameter of the key detection object displayed at the pixel position in the 720P image is determined, and the corresponding relationship between the distance value of the pixel and the average size parameter of the key detection object displayed at the pixel position in the image is set to obtain the prior frame sizes corresponding to multiple distances. Under normal circumstances, in a camera device with fixed parameters, the size of the target object at a specific distance in the generated image is corresponding. For example, in a 720P image, at any pixel position calculated to be 50m away from the camera device, the size of the pedestrian displayed in the image is (15px, 45px), and the size of the initialization prior frame can be predicted to be (15px, 45px). At any pixel position calculated to be 5m away from the camera device, the size of the pedestrian displayed in the image is (40px, 120px), and the size of the initialization prior frame can be predicted to be (40px, 120px), and so on. The above data only illustrates one implementation method, and does not represent the numerical situation in actual application, nor does it indicate that it is the best implementation method.

[0047] In one embodiment, a scene generally includes multiple types of key detection objects. In this case, the size of the prior box corresponding to the distance data may have multiple mappings.

[0048] S120, performing initialization grid division on the image acquired by the camera device, determining distance data from the midpoint of the bottom edge of the grid to the camera device, adjusting the size parameter of the initialization grid according to the size of the priori frame corresponding to the distance data to obtain an adjusted priori frame, and using the adjusted priori frame as an initialization parameter for training the target recognition model for model training;

[0049] This step will tile the prior frame on the image to be processed and wait for the algorithm adjustment of the model. The initial grid division can be performed directly in the image, but in this case, due to the excessive number of pixels, the amount of processed data will surge. Therefore, the image can be feature processed before initializing the grid division to achieve data dimensionality reduction. For example, in a 512*512 image, the feature layer of the image is obtained after feature processing, and the size is 40*40. If N prior frames are determined for each pixel, and the original image contains a total of 262144N prior frames, the prior frames determined in the feature layer are 1600N, achieving an order of magnitude of dimensionality reduction. In a more optimal implementation, the image is subjected to multiple downsampling feature processing to obtain a feature layer size of 5*5. At this time, the number of prior frames is 25N, which greatly reduces the amount of data processing.

[0050] After initial grid division in the image, a tiled initial prior frame is obtained. The initial prior frame does not take into account the distance data of its pixel points from the camera device in the actual scene. Therefore, further, based on the pixel position of the midpoint of the bottom edge of the initialized prior frame, the distance data of the pixel position from the camera device in the actual scene is determined, and then the size mapping information of the prior frame size is used to adjust the size of the initial prior frame based on the distance data.

[0051] In an example, after the 640*480 image is initialized and divided into grids, four initial prior frames are obtained, all of which are 320*240 in size, and their positions (bottom center position) are (160, 240), (480, 240), (160, 480), and (480, 480), respectively. After size adjustment, the positions of the four adjusted prior frames remain unchanged, and their sizes are adjusted to (90*30), (80*27), (120*40), and (112*37), respectively. The sizes of the above adjusted prior frames reflect the size of the target objects that may be contained in the grid location in the image. Based on this size, a more accurate prediction is achieved, reducing the number of adjustments and the amount of calculation for the prior frames.

[0052] The adjusted prior box is used as the initialization parameter for training the target recognition model for model training. For example, a series of default boxes are generated in the SSD algorithm for target detection, and the prior box is selected from the default box for training; in the FRCNN algorithm for target detection, a prior box is set, and the prior box is trained based on the IOU detection of the real box through softmax classification; YOLO2 / YOLO3 for target detection determines the size of the prior box through clustering (kmeans), and YOLO5 determines the size of the prior box through clustering (kmeans) + genetic algorithm, and the determined prior box is put into training. The present application uses the adjusted prior box as the initialization parameter for training the target recognition model for model training, and can be applied to any target recognition model that uses a prior box for model training in the prior art. Taking the YOLO5 target recognition model as an example, the calculation model determines the size of the prior frame through clustering (kmeans) + genetic algorithm. Furthermore, based on the position of the prior frame, that is, the midpoint of the bottom edge, the distance from the position to the camera device is determined, and then the size of the prior frame is adjusted according to the mapping relationship between the distance and the size of the prior frame. The adjusted prior frame is used as the initialization parameter for training the target recognition model for model training.

[0053] S130: The computing device performs abnormal condition detection on the real-time image acquired by the camera device based on the trained model.

[0054] The camera device terminal sends the image data acquired in real time to the background computing center, and the background computing center detects abnormal conditions based on the trained model.

[0055] When using the target recognition model to detect abnormal conditions, the step of initializing the prior frame is included. In a preferred embodiment, the prior frame is initialized and adjusted using the same method as step S120, that is, the image to be recognized acquired by the camera device is initialized to be grid-divided, and the distance data from the midpoint of the bottom edge of the grid to the camera device is determined. According to the size of the prior frame corresponding to the distance data, the size parameters of the initialized grid are adjusted to obtain the adjusted prior frame, and the adjusted prior frame is used as a parameter of the prediction algorithm of the target recognition model for target recognition;

[0056] In a preferred embodiment, the optimization of the second aspect is adopted. In the second aspect, a better shape of the prior frame is determined. As mentioned above, the goal of the prior frame matching is to move and adjust the prior frame in the direction of the actual frame. The closer the prior frame is to the actual frame, the fewer the number of iterations and the less the amount of calculation. Therefore, it is hoped that the prior frame is as close to the actual frame as possible. Considering that the shapes of the focus objects photographed by the camera device in different scenes are relatively uniform to a certain extent, according to this principle, the task of this step is to determine the predicted prior frame shape on each pixel. The key detection objects of anomaly detection in different application scenarios have certain shape rules. In one example, the anomaly detection method of the present invention is applied to a road traffic control scene, and its key detection objects are different types of cars. At this time, the prior frame is usually square or rectangular, and the aspect ratio of different types of cars is relatively fixed to a certain extent. According to the aspect ratio of different types of cars presented in the image, a prior frame that meets the aspect ratio parameters is set. In one example, the anomaly detection method of the present invention is applied to warehouse management and control, and its key detection objects are indoor operating tool vehicles, containers, and staff. At this time, according to the shape of the above-mentioned key detection objects in the image, a priori box that meets the aspect ratio parameters of the shape is set. In one example, the anomaly detection method of the present invention is applied to real-time monitoring scenes in parks and squares, and its key detection objects are pedestrians. At this time, the priori box is usually in human proportions, for example, the aspect ratio can be 3:1; further, when detecting special abnormal behaviors, the regular shape of the behavior presented in the image can be determined based on the specific behavior content, such as the kneeling, lying, and lying postures presented during wrestling, so as to determine the initial shape of the priori box based on the above-mentioned shape of the key detection object. The second aspect of optimization makes the priori box closer to the real image, which can reduce the amount of algorithm calculation during model training / prediction and improve the efficiency of the detection model.

[0057] In a preferred embodiment, corresponding training samples are set according to the type of object to be detected. In the case of insufficient training samples, an expansion operation is performed based on the target object calibrated in the existing training samples. The expansion operation is specifically as follows: cutting the target object based on the calibration frame, performing displacement in the training image, determining the theoretical value of the distance from the pixel at the position to the camera device based on the position of the center of the bottom edge of the calibration frame at the end point of the displacement, determining the prior frame size of the position according to the mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training frame based on the determined prior frame size. Through the above expansion operation steps, the image used for training is expanded, and different calibration information presented by the target object at different positions in the image under the same scene is generated, and the important features of the image collected by the camera device are utilized. In the case of insufficient training samples, the content of the training data is enriched, the detection accuracy is improved, and the prior art problems mentioned in the background technology are solved.

[0058] Embodiment 2: Figure 2As shown, the present invention also provides an abnormal condition detection system for a smart safety help cloud platform, comprising:

[0059] A distance determination module is used to determine a standard calibration object, move the standard calibration object so that it traverses the field of view of the camera device, and in the process of traversing the field of view, perform a step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device;

[0060] A priori frame mapping module, used to set the priori frame sizes corresponding to multiple distances according to the average size parameters of the key detection objects of the camera device;

[0061] A training module is used to initialize grid division of the image acquired by the camera device, determine the distance data from the midpoint of the bottom edge of the grid to the camera device, adjust the size parameters of the initialization grid according to the size of the prior frame corresponding to the distance data to obtain an adjusted prior frame, and use the adjusted prior frame as an initialization parameter for training the target recognition model for model training;

[0062] The detection module is a computing device that performs abnormal condition detection on the real-time image acquired by the camera device based on the trained target recognition model.

[0063] In a preferred embodiment, in the above system, the standard calibration object is moved to traverse the field of view of the camera device. During the traversal of the field of view, the distance calculation is performed for each pixel point to calculate the actual distance D between the currently traversed pixel point and the camera device:

[0064]

[0065] In the above formula (1), L is the size of the standard calibration object, f is the focal length of the camera device, and d is the image size of the standard calibration object in the camera device.

[0066] In a preferred embodiment, in the above system, a scene includes multiple types of key detection objects, and the prior box size corresponding to the distance data includes multiple mappings, each mapping is associated with each key detection object.

[0067] In a preferred embodiment, in the above-mentioned system, it is applied to real-time monitoring scenes, and its key detection objects are pedestrians. The shape of the prior frame is the human figure ratio. When abnormal behavior detection is performed, the regular shape of the behavior in the image is determined based on the specific behavior content, and the initial shape of the prior frame is determined based on the regular shape.

[0068] In a preferred embodiment, in the above system, an expansion operation is performed based on the target object calibrated in the existing training sample image. The expansion operation is specifically: cutting the target object based on the calibration frame, displacing it in the training sample image, determining the distance data from the pixel at that position to the camera device based on the position of the center of the bottom edge of the calibration frame at the end point of the displacement, determining the prior frame size of the position according to the mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training sample image based on the determined prior frame size.

[0069] The present invention provides an abnormal condition detection method and system for a smart safety help cloud platform. The method first determines the distance between the standard calibration object and the camera device in the field of view, determines the size of the prior frame corresponding to the distance according to the average size of the key detection object, adjusts the initialization grid based on the determined prior frame size, uses the adjusted prior frame as the initialization parameter for training the target recognition model for model training, and finally detects abnormal conditions on the real-time image obtained by the camera device based on the trained model. The present invention also provides a corresponding sample expansion scheme. The above method of the present invention utilizes the image characteristics collected under the hardware architecture of the smart safety help cloud platform, optimizes the prior frame of the target detection algorithm, reduces the algorithm complexity, and improves the training efficiency and prediction accuracy of the target detection model.

[0070] According to one embodiment, a program product such as a machine-readable medium is provided. The machine-readable medium may have instructions (i.e., the above-mentioned elements implemented in the form of software), which, when executed by a machine, causes the machine to perform the above-mentioned combination of various embodiments of this specification. Figure 1 Specifically, a system or device equipped with a readable storage medium may be provided, on which a software program code implementing the functions of any of the above-mentioned embodiments is stored, and a computer or processor of the system or device reads and executes the instructions stored in the readable storage medium.

[0071] In this case, the program code itself read from the machine-readable medium can realize the function of any one of the above embodiments, and thus the machine-readable code and the machine-readable storage medium storing the machine-readable code constitute part of this specification.

[0072] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code may be downloaded from a server computer or a cloud via a communication network.

[0073] Those skilled in the art should understand that the various embodiments disclosed above can be modified and altered in various ways without departing from the essence of the invention. Therefore, the protection scope of this specification should be defined by the appended claims.

[0074] It should be noted that not all steps and units in the above-mentioned processes and system structure diagrams are necessary, and some steps or units can be ignored according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above-mentioned embodiments can be a physical structure or a logical structure, that is, some units may be implemented by the same physical client, or some units may be implemented by multiple physical clients, or some components in multiple independent devices may be implemented together.

[0075] In the above embodiments, the hardware unit or module can be realized by mechanical or electrical means. For example, a hardware unit, module or processor can include permanent dedicated circuit or logic (such as special processor, FPGA or ASIC) to complete the corresponding operation. The hardware unit or processor can also include programmable logic or circuit (such as general-purpose processor or other programmable processor), which can be temporarily set by software to complete the corresponding operation. Specific implementation (mechanical method or dedicated permanent circuit or temporary circuit) can be determined based on cost and time consideration.

[0076] The specific embodiments described above in conjunction with the accompanying drawings describe exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of protection of the claims. The term "exemplary" used throughout this specification means "used as an example, instance or illustration" and does not mean "preferred" or "having advantages" over other embodiments. For the purpose of providing an understanding of the described technology, the specific embodiments include specific details. However, these technologies can be implemented without these specific details. In some instances, in order to avoid making the concepts of the described embodiments difficult to understand, well-known structures and devices are shown in block diagram form.

[0077] The above description of the present disclosure is provided to enable any person of ordinary skill in the art to implement or use the present disclosure. Various modifications to the present disclosure will be apparent to those of ordinary skill in the art, and the general principles corresponding to the present disclosure may be applied to other variations without departing from the scope of protection of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but is consistent with the widest range of principles and novel features disclosed herein.

Claims

1. A method for detecting abnormal conditions in a smart security help cloud platform, characterized in that: The method comprises the steps of: S100, determining a standard calibration object, moving the standard calibration object so that it traverses the field of view of the camera device, and in the process of traversing the field of view, executing the step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device; S110, setting a priori box sizes corresponding to a plurality of distances according to an average size parameter of an object to be detected by the camera device; S120, performing initialization grid division on the image acquired by the camera device, determining distance data from the midpoint of the bottom edge of the grid to the camera device, adjusting the size parameter of the initialization grid according to the size of the priori frame corresponding to the distance data to obtain an adjusted priori frame, and using the adjusted priori frame as an initialization parameter for training the target recognition model for model training; S130, the computing device performs abnormal condition detection on the real-time image acquired by the camera device based on the trained target recognition model.

2. The abnormal condition detection method according to claim 1, characterized in that: The standard calibration object is moved to traverse the field of view of the camera device. During the traversal of the field of view, the distance calculation is performed for each pixel point to calculate the actual distance D between the currently traversed pixel point and the camera device: Where L is the size of the standard calibration object, f is the focal length of the camera device, and d is the imaging size of the standard calibration object in the camera device.

3. The abnormal condition detection method according to claim 1, characterized in that: A scene includes multiple types of key detection objects, and the prior box size corresponding to the distance data includes multiple mappings, each of which is associated with each key detection object.

4. The abnormal condition detection method according to claim 1, characterized in that: Applied to real-time monitoring scenarios, the key detection object is pedestrians, the prior frame shape is the human figure ratio, and when abnormal situation detection is performed, the regular shape of the behavior in the image is determined based on the specific behavior content, and the initial shape of the prior frame is determined based on the regular shape.

5. The abnormal condition detection method according to claim 3, characterized in that: An expansion operation is performed based on a target object calibrated in an existing training sample image, and the expansion operation specifically includes: cutting the target object based on a calibration frame, performing displacement in the training sample image, determining distance data from the pixel at the position to the camera device based on the position of the center of the bottom edge of the calibration frame at the displacement end point, determining the prior frame size of the position according to a mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training sample image based on the determined prior frame size.

6. An abnormal situation detection system for a smart security help cloud platform, characterized in that include: A distance determination module is used to determine a standard calibration object, move the standard calibration object so that it traverses the field of view of the camera device, and in the process of traversing the field of view, perform a step of determining the distance between the standard calibration object and the camera device based on the camera device parameters, the standard calibration object parameters, and the image size of the standard calibration object in the camera device; A priori frame mapping module, used to set the priori frame sizes corresponding to multiple distances according to the average size parameters of the key detection objects of the camera device; A training module is used to initialize grid division of the image acquired by the camera device, determine the distance data from the midpoint of the bottom edge of the grid to the camera device, adjust the size parameters of the initialization grid according to the size of the prior frame corresponding to the distance data to obtain an adjusted prior frame, and use the adjusted prior frame as an initialization parameter for training the target recognition model for model training; The detection module is a computing device that performs abnormal condition detection on the real-time image acquired by the camera device based on the trained target recognition model.

7. The abnormal condition detection system according to claim 6, characterized in that: The standard calibration object is moved to traverse the field of view of the camera device. During the traversal of the field of view, the distance calculation is performed for each pixel point to calculate the actual distance D between the currently traversed pixel point and the camera device: Where L is the size of the standard calibration object, f is the focal length of the camera device, and d is the imaging size of the standard calibration object in the camera device.

8. The abnormal condition detection system according to claim 6, characterized in that: A scene includes multiple types of key detection objects, and the prior box size corresponding to the distance data includes multiple mappings, each of which is associated with each key detection object.

9. The abnormal condition detection system according to claim 6, characterized in that: Applied to real-time monitoring scenarios, the key detection object is pedestrians, the prior frame shape is the human figure ratio, and when abnormal situation detection is performed, the regular shape of the behavior in the image is determined based on the specific behavior content, and the initial shape of the prior frame is determined based on the regular shape.

10. The abnormal condition detection system according to claim 8, characterized in that: An expansion operation is performed based on a target object calibrated in an existing training sample image, and the expansion operation specifically includes: cutting the target object based on a calibration frame, performing displacement in the training sample image, determining distance data from the pixel at the position to the camera device based on the position of the center of the bottom edge of the calibration frame at the displacement end point, determining the prior frame size of the position according to a mapping relationship between the distance data and the prior frame parameters, and proportionally adjusting the calibration frame size of the target object in the training sample image based on the determined prior frame size.

Images