TW202111670A - Method and system for detecting image object - Google Patents

Abstract

The disclosure provides a method and system for detecting an image object. The method includes: receiving an abnormal alarm event; detecting a target object and an object type and an object bounding box of the target object existing in the event image; detecting a moving area range in each event image; obtaining an intersection ratio between the object bounding box of the target object and the moving area range, and accordingly finding a specific object in the target object; in response to determining that a moving object exists in the specific target object, triggering a corresponding alarm procedure based on the object type of the moving object.

Description

Image object detection method and system

The present invention relates to an image monitoring technology, and particularly relates to an image object detection method and system.

Establishing video surveillance and photography equipment and turning on the anomaly detection and alarm function is the simplest and most common way to maintain the security of important fields/parks in real time. However, in practice, there are often problems with a high false alarm rate, such as: abnormal alarm events triggered by changes in light and shadow, small animals or leaves, or mosquitoes passing in front of the lens. As image monitoring equipment is widely installed, labor costs should increase. When the manpower is limited, the real-time monitoring work will gradually be unable to load, and the abnormal warning function on the surveillance camera may be ignored or closed, and the real-time monitoring and warning capability will be lost. Therefore, for those skilled in the art, if the relevant software and hardware technology and equipment can assist the personnel to filter and accurately identify the suspicious objects that appear in the monitoring area, and then send an alarm event to notify the personnel for further confirmation and response procedures, Allows limited human resources to be used effectively.

In addition, regarding cloud AI image detection and recognition services, if video streams are uploaded to the cloud for monitoring, analysis, and control, there will be a network bandwidth bottleneck, and it is also limited by the back-end execution of image analysis. system resource. In addition, the field side also needs to deploy compatible photography equipment, and it is not easy to integrate existing built-in photography equipment, especially surveillance photography equipment that has simple abnormal image detection capabilities.

In view of this, the present invention provides an image object detection method and system, which can be used to solve the above technical problems.

The present invention provides an image object detection method, including: receiving an abnormal alarm event, wherein the abnormal alarm event includes a plurality of event images corresponding to an abnormal event; detecting at least one target object and each target existing in the aforementioned event image An object type and an object bounding box of the object; detect a moving area range in each event image based on the aforementioned event image; obtain an intersection ratio of the object bounding box and the moving area range of each target object, and based on at least one At least one specific target object is found among the target objects; in response to determining that a dynamic object exists in the at least one specific target object, a corresponding designated alarm procedure is triggered based on the object type of the dynamic object.

The invention provides an image object detection system, which includes a servo unit, an image processing unit and a processing unit. The server unit receives an abnormal alarm event, where the abnormal alarm event includes a plurality of event images corresponding to an abnormal event. The image processing unit is configured to: detect at least one target object existing in the aforementioned event image and an object type and an object bounding box of each target object; detect a moving area range in each event image based on the aforementioned event image; Obtain an intersection ratio between the object bounding box of each target object and the moving area range, and find at least one specific target object among the at least one target object accordingly. After responding to determining that a dynamic object exists in at least one specific target object, the processing unit triggers a corresponding designated alarm procedure based on the object type of the dynamic object.

Based on the above, the present invention can provide an image object detection model trained with machine learning technology to find each object category and area in the image, and then calculate the moving area based on the continuous picture, and compare the image object area with the moving area, which can be distinguished Moving objects and stationary objects. In this way, false alarms caused by changes in light and shadow and mosquitoes flying in front of the lens can be greatly reduced, and abnormal alarms can be filtered with more accurate image object detection and identification capabilities.

In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail in conjunction with the accompanying drawings.

In the embodiments of the present invention, the built-in image abnormal events of existing image surveillance devices (such as IP cameras) that are not equipped with artificial intelligence (AI) object recognition functions that have been built in the field can be used The ability to detect alarms, as the front-end computing node of edge computing, finds out the common standard transmission method that most surveillance camera equipment on the market can provide: e-mail alarms of abnormal events and image functions are used to collect information on the field side. There are abnormal events and video images of surveillance camera equipment.

In conclusion, the present invention provides an image object detection method and system, which can be implemented through a boundary computing architecture, and is more suitable for cloud AI application services. In summary, the system of the present invention can provide an image object detection model trained with machine learning technology to find out each object category and area in the image, and then calculate the moving area based on the continuous picture, and compare the image object area with the moving area , Can distinguish between moving objects and stationary objects. In this way, false alarms caused by changes in light and shadow and mosquitoes flying in front of the lens can be greatly reduced, and abnormal alarms can be filtered with more accurate image object detection and identification capabilities.

In addition, in order to maintain the detection accuracy of the machine learning image object detection model, a lot of manpower is needed to collect the model training image database and label objects. Therefore, the present invention also provides active collection of image data for which the confidence value of the detection result of the image object does not reach the predicted confidence threshold and the user report the image data for the error of the object detection result. After correction, the machine learning training is re-invested through continuously increasing fields. The domain image screen and the correction of the object image with identification defects can significantly and effectively reduce the manpower for image collection and continuously and effectively improve the detection accuracy of the object detection model. This will be further explained below.

Please refer to FIG. 1, which is a schematic diagram of a system according to an embodiment of the present invention. As shown in FIG. 1, the system 10 includes one or more image monitoring devices 100, an image object detection system 200 and one or more client devices 300.

In one embodiment, the image monitoring device 100 is, for example, a traditional webcam, a digital video recorder (DVR), a network video recorder (NVR) or other similar devices, but it is not limited to this . As shown in FIG. 1, the image monitoring device 100 may include an image event detection module 110 and an alarm module 120. The image event detection module 110 can have the capability of monitoring field images and detecting abnormal events, and can generate abnormal alarm events when an abnormal event is detected. In an embodiment, the image event detection module 110 may have one or more detection functions such as face recognition, intrusion detection, motion detection, and AI object detection, but it is not limited to this.

In addition, the alarm module 120 can provide the function of sending the above-mentioned abnormal alarm event to the image object detection system 200. In one embodiment, the above-mentioned abnormal alarm event may include an alarm message and one or more event images corresponding to the abnormal event, and the abnormal alarm event may be sent to the server unit of the image object detection system 200 via the network in the form of an email. 210.

In this embodiment, the image monitoring device 100 must be able to connect to the server unit 210 of the image object detection system 200 through the network, and obtain the sender’s email account and a set of email recipients provided by the server unit 210 account number. The sender's email can be the same as the email recipient's account, but it is not limited to this. Examples of related settings of the video surveillance device 100 in the field set by it are shown in Table 1 below, but the present invention is not limited to this. Setting bar value Sender email account site001@tl.cht.com.tw Recipient email account site001@tl.cht.com.tw Voice alert number 0999000000 SMS alert number 0999000000 The object type of the dynamic object to be detected locomotive people car Corresponding processing method for detected dynamic objects Event alarm event archive Event alarm event archive Event archive Table 1

In one embodiment, the image monitoring device 100 can use the content of Table 1 to register in the image object detection system 200, so that the image object detection system 200 can grasp the information of the image monitoring device 100 and the types of objects that need image detection. And the corresponding processing method of detected dynamic objects. For example, in the scenario in Table 1, if the object type of the detected dynamic object is a locomotive, the image object detection system 200 can correspondingly issue an event alarm and perform an event archive operation. If the object type of the detected dynamic object is a person, the image object detection system 200 can correspondingly issue an event alarm and perform an event archive operation, but the present invention is not limited to this.

In FIG. 1, the image object detection system 200 may include a server unit 210, a database unit 220, an image processing unit 230, a processing unit 240 and a detection model training unit 250. In summary, by implementing the image object detection method proposed in the present invention, the image object detection system 200 can receive an abnormal alarm event from the image monitoring device 100 in the front-end field and an event image corresponding to the abnormal event. Afterwards, the image object detection system 200 can perform image object detection, so as to perform corresponding processing after screening out the objects to be alerted. In addition, the image object detection system 200 can also provide autonomous collection of object image data that can improve machine learning training, so as to continuously refine the internal image object detection model. The following will be described in detail with FIG. 2.

Please refer to FIG. 2, which is a method for detecting an image object according to an embodiment of the present invention. The method of this embodiment can be executed by the image object detection system 200 of FIG. 1. The details of each step in FIG. 2 are described below with the content of FIG. 1.

First, in step S210, the servo unit 210 may receive an abnormal alarm event. For example, the server unit 210 may receive an email containing an abnormal alarm event and one or more related event images from the image monitoring device 100. In one embodiment, the above-mentioned abnormal alarm event and related event images may be stored in the database unit 220, but it is not limited thereto.

Please refer to FIG. 3, which is a schematic diagram of an abnormal alarm event drawn according to an embodiment of the present invention. In this embodiment, the abnormal alarm event 30 may include, for example, the image monitoring device 100 and related information about the abnormal event (such as equipment number, name, occurrence time and type of the abnormal event) and event images 311 to 313, but the present invention is not limited to this.

Then, in step S220, the image processing unit 230 can detect the target object existing in the aforementioned event image and the object type and object bounding box of each target object. As shown in FIG. 1, the image processing unit 230 may include an image object detection and classification module 232 and a movement area calculation module 234. The image object detection and classification module 232 can detect the presence of the event image according to the content of Table 1, for example. People, locomotives and cars in, and the corresponding object bounding box can frame the detected target object.

Please refer to FIG. 4, which is a schematic diagram of the event image marked with the bounding box of the object shown in FIG. 3. As shown in FIG. 4, in the event image 311, the image object detection and classification module 232 may, for example, mark multiple target objects 711 to 715 with corresponding object bounding boxes; in the event image 312, The image object detection and classification module 232 may, for example, mark multiple target objects 721 to 727 with the corresponding object bounding box; in the event image 313, the image object detection and classification module 232 may, for example, detect After multiple target objects 731 to 735 are detected, they are marked with corresponding object bounding boxes, but the present invention is not limited to this.

In one embodiment, the image object detection and classification module 232 can input the event images 311 to 313 into the first image object detection model to find the target object and the object type of each target object from the first image object detection model And the bounding box of the object. The above-mentioned first image object detection model is, for example, a machine learning model trained based on a model training image database, but it is not limited to this.

In one embodiment, the first image object detection model may include a plurality of sub-image object detection models connected in series, and these sub-image object detection models can be used to detect target objects corresponding to different object types. Take the content of Table 1 as an example. In order to detect target objects belonging to human, motorcycle, automobile and other object types, the above-mentioned first image object detection model can include 3 sub-image object detection models, and they can be used to find out which objects belong to Target objects such as people, motorcycles, automobiles, etc., can improve the accuracy of each target object, but the present invention is not limited to this.

In addition, in one embodiment, the above-mentioned first image object detection model can also generate a detection confidence value for each target object. If the first image object detection model has a higher detection confidence value for a certain target object, it means that the first image object detection model more determines the object type of the target object, and vice versa.

In one embodiment, the object type, object bounding box, and detection confidence value of each target object in each event image 311 to 313 can be illustrated in Table 2 below. Event image Target object Object type Object bounding box Detection confidence value 311 711 people (X ₇₁₁ ,Y ₇₁₁ ,W ₇₁₁ ,H ₇₁₁ ) 95% 311 712 locomotive (X ₇₁₂ ,Y ₇₁₂ ,W ₇₁₂ ,H ₇₁₂ ) 95% 311 713 people (X ₇₁₃ ,Y ₇₁₃ ,W ₇₁₃ ,H ₇₁₃ ) 98% 311 714 locomotive (X ₇₁₄ ,Y ₇₁₄ ,W ₇₁₄ ,H ₇₁₄ ) 30% 311 715 locomotive (X ₇₁₅ ,Y ₇₁₅ ,W ₇₁₅ ,H ₇₁₅ ) 28% 312 721 people (X ₇₂₁ ,Y ₇₂₁ ,W ₇₂₁ ,H ₇₂₁ ) 92% 312 722 locomotive (X ₇₂₂ ,Y ₇₂₂ ,W ₇₂₂ ,H ₇₂₂ ) 88% 312 723 people (X ₇₂₃ ,Y ₇₂₃ ,W ₇₂₃ ,H ₇₂₃ ) 96% 312 724 locomotive (X ₇₂₄ ,Y ₇₂₄ ,W ₇₂₄ ,H ₇₂₄ ) 30% 312 725 locomotive (X ₇₂₅ ,Y ₇₂₅ ,W ₇₂₅ ,H ₇₂₅ ) 28% 312 726 locomotive (X ₇₂₆ ,Y ₇₂₆ ,W ₇₂₆ ,H ₇₂₆ ) 70% 312 727 people (X ₇₂₇ ,Y ₇₂₇ ,W ₇₂₇ ,H ₇₂₇ ) 65% 313 731 people (X ₇₃₁ ,Y ₇₃₁ ,W ₇₃₁ ,H ₇₃₁ ) 85% 313 732 locomotive (X ₇₃₂ ,Y ₇₃₂ ,W ₇₃₂ ,H ₇₃₂ ) 95% 313 733 people (X ₇₃₃ ,Y ₇₃₃ ,W ₇₃₃ ,H ₇₃₃ ) 95% 313 734 locomotive (X ₇₃₄ ,Y ₇₃₄ ,W ₇₃₄ ,H ₇₃₄ ) 30% 313 735 locomotive (X ₇₃₅ ,Y ₇₃₅ ,W ₇₃₅ ,H ₇₃₅ ) 28% Table 2

In Table 1, the object bounding box of the target object 711 can be expressed as (X ₇₁₁ , Y ₇₁₁ , W ₇₁₁ , H ₇₁₁ ), where X ₇₁₁ and Y ₇₁₁ are the X coordinate and Y of a reference point of the object bounding box, respectively Coordinates, and W ₇₁₁ and H ₇₁₁ respectively define the width and length of the object's bounding box. Based on this, a person with ordinary knowledge in the field should be able to understand the representation of the object bounding box of other target objects accordingly, which will not be repeated here.

In some embodiments, if there is only one event image received, the image object detection system 200 can directly perform corresponding operations based on the contents of Table 1 after finding the target object, such as sending event alarms and event notifications. And/or event archive, etc., but not limited to this.

After step S230, the movement area calculation module 234 can perform step S240 to detect the movement area range in each event image 311-313 based on the event images 311-313. In different embodiments, the moving area calculation module 234 can perform, for example, a continuous image subtraction (Temporal Differencing) technique or a background subtraction technique on the event images 311 to 313 to obtain the event images 311 to 313. The range of the moving area. In the embodiment of the present invention, the moving area range can be roughly understood as a range in which an object is moving, but the present invention may not be limited to this.

Please refer to FIG. 5, which is a schematic diagram of an event image based on the marked moving area shown in FIG. 3. As shown in FIG. 5, the movement area calculation module 234 can find the movement area ranges 811 and 812 in the event image 311, the movement area ranges 821 and 822 in the event image 312, and the movement in the event image 313. Area scope 813 and 813. And, as can be seen from FIG. 5, the moving area ranges 811, 821, and 831 are the same as each other (hereinafter referred to as the moving area 1), and the moving area ranges 812, 822, and 832 are the same as each other (hereinafter referred to as the moving area 2). The invention may not be limited to this.

In an embodiment, the movement areas 1 and 2 can be integrated into Table 3 below. Moving area range Moving area 1 811=821=831=>(X ₈₀₁ ,Y ₈₀₁ ,W ₈₀₁ ,H ₈₀₁ ) Moving area 2 812=822=832 =>(X ₈₀₂ ,Y ₈₀₂ ,W ₈₀₂ ,H ₈₀₂ ) table 3

In addition, it should be understood that although step S240 is shown after step S230 in FIG. 2, it is not used to limit the execution order. In some embodiments, the execution order of the two steps can also be reversed, but the present invention is not limited to this.

After that, in step S240, the moving area calculation module 234 can obtain the intersection ratio of the object bounding box of each target object 711~735 and the moving area range 811~832, and then find a specific target among the target objects 711~735 object.

In one embodiment, for the first target object among the target objects 711 to 735, the moving area calculation module 234 can obtain an intersection area between the object bounding box of the first target object and the moving area range. After that, the moving area calculation module 234 can divide the intersection area by the object bounding box of the first target object to obtain the intersection ratio between the object bounding box of the first target object and the moving area range.

Please refer to FIG. 6, which is a schematic diagram of obtaining the intersection ratio between the bounding box of the object and the range of the moving area according to FIGS. 4 and 5. Taking the target object 711 in the event image 311 as an example, when the moving area calculation module 234 obtains the intersection ratio between the object bounding box of the target object 711 and the moving range area 811, the moving area calculation module 234 may first obtain the target The intersection area of the object bounding box of the object 711 and the moving area range 811. In the scenario of FIG. 6, this intersection area can be regarded as the entire object bounding box of the target object 711. Then, the moving area calculation module 234 can divide the intersection area by the object bounding box of the target object 711 to obtain the intersection ratio (ie, 100%) of the object bounding box of the target object 711 and the moving area range 811.

Taking the target object 723 in the event image 312 as an example, when the moving area calculation module 234 obtains the intersection ratio between the object bounding box of the target object 723 and the moving range area 822, the moving area calculation module 234 may first obtain The intersection area of the object bounding box of the target object 723 and the moving area range 822. In the scenario of FIG. 6, this intersection area can be regarded as the entire object bounding box of the target object 723. After that, the moving area calculation module 234 can divide the intersection area by the object bounding box of the target object 723 to obtain the intersection ratio (ie, 100%) of the object bounding box of the target object 723 and the moving area range 822.

As for the intersection ratio between the object bounding box of other target objects and the corresponding moving range area, a person with ordinary knowledge in the art should be able to deduce based on the above teachings, which will not be repeated here.

After that, the moving area calculation module 234 can find the specific target object in each target object 711-735 according to the intersection ratio corresponding to it. In this embodiment, the intersection rate of the object bounding box of each specific target object and the corresponding moving area range is higher than the preset intersection rate threshold. In short, the moving area calculation module 234 can find one or more of the target objects 711 to 735 with a higher intersection rate as the specific target object, but the present invention is not limited to this.

After that, the processing unit 240 may classify the target objects to find dynamic objects and/or static objects among them. In one embodiment, the classification result can be illustrated in Table 4 below. Dynamic objects in specific target objects Static objects with low intersection ratio Moving area 1 711,712, 721,722,726,727, 731,732,733 Moving area 2 713, 723 Not in the moving area 714,715, 724,725, 734,735 Table 4

Afterwards, in response to determining that there is a dynamic object in the specific target object, the detected object corresponding processing module 242 of the processing unit 240 can trigger a corresponding designated alarm procedure based on the object type of the dynamic object. For example, it can be seen from Table 4 that the object types to which the dynamic objects belonged are locomotives and people, which are consistent with the content of Table 1, so the processing unit 240 can issue event alarms accordingly and perform event archiving operations.

In an embodiment, if there are only static objects in the target object that do not intersect with any moving area range, the detected object corresponding processing module 242 can ignore the above abnormal alarm event accordingly.

In one embodiment, the image object detection system 200 of the present invention can also collect and correct the inaccurate image data of the object detection result, and refine the image object detection model by machine learning technology. The accuracy test threshold is to trigger the system to replace the new model. The detailed description is as follows.

In one embodiment, the image object training image data collection module 244 in the processing unit 240 can be used to collect the object to be confirmed and add the object to be confirmed to the model training image database 254 in the detection model training unit 250. In different embodiments, the above-mentioned object to be confirmed may have one or more of the following aspects according to requirements.

For example, in one embodiment, the image object training image data collection module 244 can find from the target objects 711 to 735 that the detection confidence value in each event image 311 to 313 is less than the confidence threshold (for example, 50 One or more of %) as the object to be confirmed. Based on the content of Table 2, the target objects 714, 715, 724, 725, 734, and 735 can be defined as objects to be confirmed, but the present invention is not limited to this.

For another example, in the continuous first event image and the second event image, the image object training image data collection module 244 can obtain a first target object and the first target object in the first event image. The type of object. Thereafter, the image object training image data collection module 244 can determine whether there is a second target object corresponding to the first target object in the second event image. In response to the presence of the second target object in the second event image, the image object training image data collection module 244 can obtain the second object type of the second target object. Afterwards, reflecting that the second object type is different from the first object type, the image object training image data collection module 244 can determine that at least one of the first target object and the second target object belongs to the object to be confirmed. In short, for the same target object in different event images, if the target object has different object types detected in different event images, the image object training image data collection module 244 can use this target object Treated as an object to be confirmed.

In addition, the image object training image data collection module 244 can also directly use the target object selected by the user as the object to be confirmed. Specifically, as shown in FIG. 2, the client device 300 may include a receiving and feedback module 310. In one embodiment, the receiving and feedback module 310 is capable of receiving one or more alarm messages sent by the detection object corresponding processing module 242 of the processing unit 240, where the alarm messages may include: voice, text message, video screen, Program instructions, driver hardware procedures, etc. When the receiving and feedback module 310 receives emails, web browsers, or applications, it can link to whether the detection result of the target object is accurate based on one or more abnormal event image screens and the message of the detected alarm object Or button or image, the reply information will be returned to the image object training image data collection module 244, and the image object training image data collection module 244 will return one or more inaccurate alarm information images and target object frame selection information storage In the database unit 220, after the personnel correctly label the data, they are added to the model training map database 254.

After obtaining the object to be confirmed, the image object detection model training module 252 in the detection model training unit 250 can use machine learning technology to train the second image object detection model using the model training image database 254, and obtain the second image object detection model. The detection accuracy of the image object detection model. In response to the detection accuracy of the second image object detection model being higher than the model check threshold, the image object detection model training module 252 can replace the first image object detection model with the second image object detection model. That is, if another abnormal event alarm occurs in the future, the image processing unit 230 can use the second image object detection model to detect the target object in the related event image, but the invention is not limited to this.

In summary, the method and system of the present invention can provide a cloud AI application service, using a boundary computing architecture, using the anomaly detection function of the field-side monitor device as the boundary endpoint, and using the commonly built-in electronics of the image monitoring device The function of the email alarm event image screen is used as a service integration interface to enable existing image surveillance equipment that does not have the AI object recognition function to have AI object recognition and filtering capabilities.

The present invention can provide only one or more continuous images of event images, use machine learning object detection models to detect image objects and calculate scene image differences, and compare the content of the intersection area between the image object area and the moving area, thereby Distinguish between moving objects and stationary objects, and perform alarms and specified procedures for preset specific objects. It can significantly reduce the misreporting incidents that are criticized, and greatly reduce the labor processing cost of the false alarm incidents of the image monitoring equipment.

The present invention can provide the user to pre-set the types of filtered objects. When image object detection is performed through the machine learning object detection model, the object detection will use one or more object detections that meet the required detection object type Model, cascade to form a cascaded object detection model, which effectively improves the accuracy of image object detection.

The present invention can filter out the moving objects that need to be alerted in the images of each monitoring field according to the user's setting, and drive the corresponding processing program. The object of the processing program also includes alarm notification to the user's local call and smart mobile device or activation of a specified in-device program, and can also be used as an interactive way to turn on a specific device.

The present invention can provide a training image data collection method that effectively refines the object detection model: by analyzing the object detection results of each alarm image, collecting object image information with a detected object confidence value lower than a preset accurate threshold, and collecting continuous images The image data of the same object area of the object detection result but the object category name is different, as well as the information feedback after the alert user is collected, can reduce the cost of image data collection; the image data with poor accuracy can be correctly labeled and included in the machine The training image data collection of learning can continuously and effectively improve the prediction accuracy of the image object detection model.

The present invention can be used as a solution for upgrading the image monitoring equipment without AI object judgment capability to protection with the AI object judgment function on the premise that the image surveillance equipment does not need to be fully updated.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the present invention shall be determined by the scope of the attached patent application.

10: System 100: Video surveillance device 110: Video event detection module 120: Alarm module 200: Image object detection system 210: servo unit 220: database unit 230: image processing unit 232: Image object detection and classification module 234: Mobile area calculation module 240: processing unit 242: Detected object corresponding processing module 244: Image object training image data collection module 250: Detection model training unit 252: Image object detection model training module 254: Model training image database 30: Abnormal alarm event 300: client device 311~313: Event video 711~735: target object 811, 812: moving area range S210~S250: steps

Fig. 1 is a schematic diagram of a system according to an embodiment of the present invention. FIG. 2 shows a method for detecting an image object according to an embodiment of the present invention. Fig. 3 is a schematic diagram of an abnormal alarm event drawn according to an embodiment of the present invention. FIG. 4 is a schematic diagram of the event image marked with the bounding box of the object shown in FIG. 3. FIG. 5 is a schematic diagram of an event image according to the marked moving area shown in FIG. 3. FIG. 6 is a schematic diagram of obtaining the intersection ratio between the bounding box of the object and the range of the moving area according to FIG. 4 and FIG. 5.

S210~S250: steps

Claims (13)

An image object detection method, including: Receiving an abnormal alarm event, where the abnormal alarm event includes multiple event images corresponding to an abnormal event; Detecting at least one target object existing in the event images and an object type and an object bounding box of each target object; Detecting a moving area range in each event image based on the event images; Obtain an intersection ratio of the object bounding box of each target object and the moving area range, and find at least one specific target object among the at least one target object accordingly; In response to determining that a dynamic object exists in the at least one specific target object, a corresponding designated alarm procedure is triggered based on the object type of the dynamic object. The method described in item 1 of the scope of patent application, wherein the abnormal alarm event comes from an image monitoring device. The method described in item 1 of the scope of patent application, wherein the intersection rate of the object bounding box of each specific target object and the moving area range is higher than a preset intersection rate threshold. According to the method described in claim 1, wherein the step of detecting the range of the moving area in the specific field based on the event images includes: Perform a continuous image subtraction technique or a background subtraction technique on the event images to obtain the moving area range in each of the event images. For the method described in item 1 of the scope of patent application, the step of obtaining the intersection ratio of the object bounding box of each target object and the moving area range includes: For a first target object among the at least one target object, obtaining an intersection area of the object bounding box of the first target object and the moving area range; The intersection area is divided by the object bounding box of the first target object to obtain the intersection ratio of the object bounding box of the first target object and the moving area range. According to the method described in item 1 of the scope of the patent application, wherein it is reflected in the at least one target object that there is only a static object that does not intersect the range of the moving area, the method further includes ignoring the abnormal alarm event. According to the method described in claim 1, wherein the step of detecting the at least one target object and the object type and the object bounding box of each of the target objects existing in the event images includes: The event images are input into a first image object detection model to find the at least one target object and the object type and the object bounding box of each target object by the first image object detection model. According to the method described in claim 7, wherein the first image object detection model includes a plurality of sub-image object detection models connected in series, and the sub-image object detection models are used to detect objects corresponding to different objects. The at least one target object of the type. As the method described in item 7 of the scope of patent application, it also includes: Collect at least one object to be confirmed, and add the at least one object to be confirmed into a model training image database; Using a machine learning technology to train a second image object detection model using the model training image database, and obtain a detection accuracy rate of the second image object detection model; Reflecting that a detection accuracy rate of the second image object detection model is higher than a model check threshold, the first image object detection model is replaced by the second image object detection model. For the method described in item 9 of the scope of patent application, the step of collecting the at least one object to be confirmed includes: Obtain a detection confidence value when detecting each target object in each event image from the first image object detection model; Find the at least one object to be confirmed among the at least one target object, wherein the detection confidence value of each object to be confirmed corresponding to each event image is less than a confidence value threshold. According to the method described in claim 9, wherein the event images include a first event image and a second event image consecutively, and the step of collecting the at least one object to be confirmed includes: Acquiring a first target object in the first event image and a first object type of the first target object; Determine whether there is a second target object corresponding to the first target object in the second event image; In response to the presence of the second target object in the second event image, obtaining a second object type of the second target object; In response to the second object type being different from the first object type, it is determined that at least one of the first target object and the second target object belongs to the at least one object to be confirmed. For the method described in item 9 of the scope of patent application, the step of collecting the at least one object to be confirmed includes: One or more objects selected by a user from the at least one target object are used as the object to be confirmed. An image object detection system, including: A server unit that receives an abnormal alarm event, wherein the abnormal alarm event includes a plurality of event images corresponding to an abnormal event; An image processing unit configured to: Detecting at least one target object existing in the event images and an object type and an object bounding box of each target object; Detecting a moving area range in each event image based on the event images; Obtain an intersection ratio of the object bounding box of each target object and the moving area range, and find at least one specific target object among the at least one target object accordingly; A processing unit that, after responding to determining that a dynamic object exists in the at least one specific target object, triggers a corresponding designated alarm procedure based on the object type of the dynamic object.

Images