A method and system for detecting events in a monitored video image in conjunction with a heterogeneous electronic fence

Abstract

The application discloses a kind of monitoring video image event detection methods combined with heterogeneous electronic fence.The application utilizes coordinate conversion technology to dynamically build video electronic fence into geographic supervision graph spot, and aligns the converted video electronic fence with monitoring video image, filters out the image area that user is really interested in in the process of Mask R-CNN algorithm, greatly improves the recognition efficiency and accuracy of video image.The application can not only be applied to event detection of natural resource supervision, but also can be applied to monitoring scenes of other various high point cameras.

Description

Technical Field

[0001] This invention relates to the field of image event detection, and more specifically to a method for detecting surveillance video image events in conjunction with a heterogeneous electronic fence. Background Technology

[0002] Currently, typical electronic fence event detection relies almost entirely on geographic electronic fences, which are often difficult to obtain. Furthermore, the electronic fences integrated into surveillance cameras are usually fixed video fences, making them unsuitable for event detection in dynamically monitored areas. Existing event detection methods either depend on the geographic information of the event's occurrence or cannot implement dynamic electronic fences, resulting in an inability to dynamically and accurately monitor natural resource monitoring targets in real time. Summary of the Invention

[0003] To overcome the shortcomings and deficiencies of existing technologies, geographic electronic fences are transformed into video electronic fences through geographic video coordinate transformation technology, and the Mask R-CNN algorithm is used to perform real-time detection on the patrol images of the surveillance video cameras, thereby achieving real-time and accurate monitoring of natural resources.

[0004] The technical solution of this invention is a method for detecting events in surveillance video images combined with heterogeneous electronic fences, comprising the following steps:

[0005] Step 1: By configuring the camera patrol strategy, real-time monitoring video images are collected to obtain patrol image information;

[0006] Step 2: Calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel;

[0007] Step 3: Based on the information from the patrol images and the internal and external parameters of the camera, convert the land parcels that need to be monitored into video electronic fences;

[0008] Step 4: Use the Mask R-CNN network to perform event detection on the cruise images. The specific processing procedure of the Mask R-CNN network is as follows: the input image extracts a shared feature map through the backbone network composed of the Residual Network (ResNet) and the Feature Pyramid Network (FPN). This feature map is then input into the Region Proposal Network (RPN) to obtain candidate bounding box regions. The candidate bounding box regions are filtered based on the video geofence in Step 3. The filtered regions are then input into the Region of Interest (ROI) matching layer (RoIAlign). RoIAlign uses bilinear interpolation to replace the quantization operation in the feature aggregation process. Finally, the ROI is input into the mask branch and the FAST-RCNN branch respectively. The final network outputs the predicted mask bounding box vertex coordinates, the predicted target bounding box center point coordinates and length and width data, and the predicted target bounding box category.

[0009] Furthermore, the cruise image information in step 1 is V ptz ,in:

[0010] p∈[0, 360]

[0011] t∈[-180, 180]

[0012] z∈[1, 50]

[0013] p represents the range of horizontal movement of the camera; t represents the range of vertical movement of the camera; z represents the zoom range.

[0014] Furthermore, the land parcels in step 2 are represented as follows:

[0015] O m =MultiPolygon m (((Lng n Lat n )))

[0016] m∈[1,M]

[0017] n∈[1, N]

[0018] Where M represents the number of monitored land parcels within the camera's field of view, and N represents the number of polygon sides corresponding to the monitored land parcels; Lng n and Lat n MultiPolygon represents the latitude and longitude coordinates of the vertices of a polygon. m (((Lng n Lat n ))) is a way of representing polygons.

[0019] Furthermore, in step 3, a mapping model between geographic coordinates and video images under varying camera postures is first constructed. Then, combining the camera parameters captured in the images, the mapping model is used to convert the land parcels into video fences. The representation of the video electronic fence is as follows:

[0020] o m =MultiPolygon m (((u n vn)))

[0021] m∈[1,M]

[0022] n∈[1, N]

[0023] Among them, u n and v nThis represents the pixel coordinates of the video image under the pan-tilt-z (PTZ) mode of the camera corresponding to the polygon, where p represents the range of horizontal movement of the camera; t represents the range of vertical movement of the camera; z represents the zoom range; M represents the number of monitored land parcels within the camera's field of view; and N represents the number of polygon sides corresponding to the monitored land parcels. MultiPolygon m (((Lng n Lat n ))) is a way of representing polygons.

[0024] Furthermore, in the actual conversion process, u is filtered out. n and v n The occurrence of negative values.

[0025] Furthermore, the input image in step 4 is the cruise image obtained in step 1; after obtaining the candidate box region, the region with high score is calculated using the non-maximum suppression (NMS) algorithm, and regions with high overlap, exceeding the image boundary, and low confidence are removed.

[0026] Furthermore, in step 4, the video electronic fence is proportionally converted into a fence with corresponding feature size, and then the intersection of the fence with the candidate box area with the corresponding feature size is calculated. The area with an intersection area greater than a certain threshold is retained to complete the area filtering.

[0027] Furthermore, RoIAlign's processing flow is as follows: First, it iterates through the selected candidate regions while keeping the floating-point numbers unchanged, divides the candidate regions into k*k units, calculates four fixed coordinate values ​​in each unit, calculates the values ​​of these four positions using bilinear interpolation, cancels the quantization process, and uses max pooling to aggregate the position values ​​to obtain the region of interest.

[0028] Furthermore, the mask branch includes deconvolutional and convolutional layers; the FAST-RCNN branch has fully connected layers, as well as bounding box classification and location regression operations, using the following loss function L:

[0029] L = L cls +L box +L mask

[0030] Among them, the border segmentation error L mask This refers to the error in the segmentation mask, and the error in the border position L. box Calculate the positional error and classification error L of the predicted bounding box output by the Mask R-CNN network. cls Calculate the error in the predicted object category output of the Mask R-CNN network.

[0031] This invention also provides a surveillance video image event detection system combined with a heterogeneous electronic fence, comprising the following modules:

[0032] The image acquisition module is used to acquire real-time monitoring video images by configuring the camera cruise strategy and obtain cruise image information;

[0033] The land parcel calculation module is used to calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel.

[0034] The video electronic fence acquisition module is used to convert the land parcels that need to be monitored into video electronic fences based on the information from the patrol images and the internal and external parameters of the camera.

[0035] The event detection module is used to perform event detection on cruise images using the Mask R-CNN network. The specific processing procedure of the Mask R-CNN network is as follows: the input image extracts a shared feature map through a backbone network composed of a ResNet residual network and a Feature Pyramid Network (FPN). This feature map is then input into the Region Proposal Network (RPN) to obtain candidate bounding box regions. The candidate bounding box regions are filtered based on video geofencing. The filtered regions are then input into the Region of Interest (ROI) matching layer (RoIAlign). RoIAlign uses bilinear interpolation to replace the quantization operation in the feature aggregation process. Finally, the ROI is input into the mask branch and the FAST-RCNN branch respectively. The final network outputs the predicted mask bounding box vertex coordinates, the predicted target bounding box center point coordinates and length and width data, and the predicted target bounding box category.

[0036] The monitoring module is used to merge and store the event results detected by video images, the corresponding original video images, and video camera information data into the database, and then transmit them to the display module to facilitate users' real-time and accurate monitoring of natural resources.

[0037] This invention utilizes coordinate transformation technology to dynamically construct video electronic fences from geographic monitoring patches, and aligns the transformed video electronic fences with the monitoring video images. During the Mask R-CNN algorithm process, it filters out the image regions of true interest to the user, significantly improving the efficiency and accuracy of video image recognition. This invention can be applied not only to event detection in natural resource monitoring but also to various other high-point camera monitoring scenarios. Attached Figure Description

[0038] Figure 1 This is a flowchart of the overall process of the method of the present invention.

[0039] Figure 2 This is a flowchart of the Mask R-CNN network according to an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the event detection results in an embodiment of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0042] This invention proposes a method for event detection in surveillance video images combined with heterogeneous electronic fences. Surveillance images from surveillance cameras are collected in real time, while data on natural resource monitoring land parcels is obtained from government agencies such as the Natural Resources and Planning Bureau. The user-provided monitoring land parcel data needs to be processed by professional software such as ArcMap to be transformed into structured data. The central server needs to collect all images within the field of view of each video camera and calculate the monitored land parcels for each camera based on the positional relationship between the cameras and the monitored land parcels. Event detection refers to identifying vehicles (excavators, bulldozers, tractors, and other agricultural machinery), personnel, and terrain features (pits, water bodies, sand piles, roads, buildings, and building materials) in the images.

[0043] like Figure 1 As shown, the technical solution of the present invention is a real-time and accurate event detection method combining heterogeneous electronic fences, comprising the following steps:

[0044] Step 1: Configuring a camera patrol strategy to acquire real-time monitoring video images, obtaining patrol image information, and transmitting it to the central server; the image information is V. ptz .

[0045] p∈[0, 360]

[0046] t∈[-180, 180]

[0047] z∈[1, 50]

[0048] Where p represents the horizontal movement of the camera, i.e., the camera's rotation; t represents the vertical movement, i.e., the camera lens's pitch; and z represents zoom, i.e., adjusting the camera's focal length. Then V ptz This indicates the patrol images taken by the video camera under the PTZ (Pan-Touch) mode. Currently, most cameras support panoramic patrol or preset point patrol, so a patrol strategy can be configured to capture all monitoring images within the field of view of the tower video camera.

[0049] Step 2: Calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel.

[0050] O m =MultiPolygon m (((Lng n Lat n )))

[0051] m∈[1,M]

[0052] n∈[1, N]

[0053] Where M represents the number of monitored land parcels within the camera's field of view, and N represents the number of polygon sides corresponding to the monitored land parcels; Lng n and Lat n This represents the latitude and longitude coordinates of the vertices of the polygon, while

[0054] MultiPolygon m (((Lng n Lat n ))) is a way of representing polygons, generally known as WKT format.

[0055] Step 3: Based on the PTZ information of the patrol images and the internal and external parameters of the cameras, convert the land parcels that need to be monitored into video electronic fences.

[0056] While video image data can realistically display spatial information, it cannot be directly and deeply integrated with geographic information systems (GIS) due to the lack of geographic attributes. Generally, video image data does not possess geospatial coordinates. To endow video image data with geospatial attributes and maintain the consistency of reference coordinates, it is usually necessary to associate these image data with a coordinate system and perform coordinate transformations.

[0057] Construct a mapping model between geographic coordinates and video images under camera device posture changes:

[0058]

[0059] Among them, f x f y d is the camera focal length; x d y Let be the physical pixel dimensions of the camera sensor in the horizontal and vertical directions; u0 and v0 are the principal point coordinates of the image pixels; K is the intrinsic parameter matrix determined only by parameters related to the internal structure of the camera; [R|t] is the extrinsic parameter matrix determined by the rotation matrix R and translation vector t of the camera relative to the world coordinate system; (X w Y w Z w The coordinates (X, Y) represent the true position coordinates of an object in the world coordinate system, which can be converted to camera coordinates (X, Y) using the rotation matrix R and translation vector t. c Y c Z c Therefore, the above formula represents the coordinates (X) in the camera coordinate system. c Y c Z c Image pixel coordinates (u, v) and real position coordinates in the world coordinate system (X). w Yw Z w ) relationship.

[0060] Based on the monitored land parcel data described in step 2, combined with the camera PETZ parameters from image capture, the coordinate transformation technique described above is used to convert it into a video fence:

[0061] O m =MultiPolygon m (((u n v n )))

[0062] m∈[1,M]

[0063] n∈[1, N]

[0064] Among them, u n and v n This represents the pixel coordinates of the video image under the PTZ (ptz) of the camera corresponding to the polygon. In the actual conversion process, u... n and v n Negative values ​​may occur, which correspond to electronic fences in areas outside the field of view of the camera pan-tilt unit (PTZ). Therefore, these fences can be filtered out.

[0065] Step 4: Use the Mask R-CNN network to perform event detection on the cruise images. Mask R-CNN is a very flexible framework that is derived from Faster R-CNN and the fully convolutional network FCN, but with many improvements. Mask R-CNN can perform various image processing tasks such as object classification, object detection, semantic segmentation, and instance segmentation.

[0066] To achieve fast and accurate image recognition, the Mask R-CNN network employs the classic object detection network Faster R-CNN and the classic semantic segmentation network FCN. Faster R-CNN performs object detection quickly and accurately; FCN performs semantic segmentation precisely. Both algorithms are classics in their respective fields. Mask R-CNN is more complex than Faster R-CNN, but it still achieves a speed of 5fps, comparable to the original Faster R-CNN. Due to the discovery of pixel bias issues in ROI Pooling, a corresponding ROI Align strategy was proposed. Combined with the precise pixel masking of FCN, it achieves high accuracy. The overall idea of ​​the Mask R-CNN algorithm is simple: it adds FCN to the original Faster R-CNN algorithm to generate the corresponding mask branch, i.e., RPN + RoIAlign + Fast-RCNN + FCN (see appendix). Figure 2.

[0067] To achieve pixel-level image segmentation, the detailed implementation steps are as follows:

[0068] (1) Input the sample data to be detected, i.e., video image V ptz Feature maps are extracted using residual convolutional networks and feature pyramid networks.

[0069] (2) The extracted feature map is input into the RPN (Region Proposal Network). The core operation of RPN is to generate a series of anchor boxes according to a certain rule based on the feature map. Assuming the size of the feature map is W×H, then W×H×k anchor boxes are generated, that is, k anchor boxes are generated for each point on the feature map. Generally, k is 9, that is, 9 anchor boxes are selected. These 9 anchor boxes have three shapes with aspect ratios of {1∶1, 1∶2, 2∶1} and three area scales, which correspond to the generation of anchor boxes. A large number of anchor boxes are generated according to the above operation. After ignoring the anchor boxes that cross the boundary, there are generally still thousands to tens of thousands. There is a lot of overlap between the candidate boxes corresponding to these anchor boxes. RPN uses the non-maximum suppression method (NMS algorithm) to remove regions with high overlap rate, regions that cross the image boundary and regions with low confidence, and calculates the regions with high scores.

[0070] (3) Since the PTZ of the geofence converted to a video fence in step 3 is consistent with that of the cruise image, their geographical locations remain the same. Because the aforementioned area suggestion network is also described based on pixel coordinates, it can be implemented using the MultiPolygon video geofence described in step 3. m (((u n v n The generated regions are then further filtered by the network. Assuming the feature map size is W×H, the MultiPolygon... m (((u n v n By proportionally converting the fence to the corresponding feature size, and determining whether the two overlap, the region selection can be completed.

[0071] (4) RoI Align (Region of Interest Alignment) is an important component of the Mask R-CNN algorithm. To better detect small targets, RoI Pooling is changed to RoI Align. RoI Align cancels the quantization operation. First, it traverses the previously selected candidate regions while keeping the floating-point numbers unchanged. The candidate regions are divided into k*k units. In each unit, four fixed coordinate values ​​are calculated. The values ​​of these four positions are calculated using the bilinear interpolation method. The quantization operation is canceled. The position values ​​are summed using the max pooling method to obtain the region of interest (RoI).

[0072] (5) Input the above region of interest (RoI) into the mask branch, and obtain a 14*14*256 feature map through the deconvolution layer. Then, output a 28*28*80 mask through the 2*2*80 convolution layer.

[0073] Simultaneously, the aforementioned Region of Interest (RoI) is input into the FAST-RCNN network, passing through two 1024-dimensional fully connected layers. The output 1024-dimensional feature vector is used for bounding box classification and location regression. Finally, the network outputs the predicted mask box vertex coordinates, the predicted target box center point coordinates and length and width data, and the predicted target box category.

[0074] (6) Loss function calculation:

[0075] L = L cls +L box +L mask

[0076] Among them, the border segmentation error L mask It is the error of the segmentation mask; the error of the border position L box Calculate the positional error of the predicted bounding box output by the network; classification error L cls Calculate the error in the predicted target category of the network output.

[0077] The above steps can detect objects in video images (including image information such as the outline of objects within the attached electronic fence, the object's category, and coordinates). Based on the detection results, it can be determined whether a target object, such as a vehicle, pit, or building, appears in the video image. (See attached image.) Figure 3 .

[0078] Step 5: Merge the results of video image detection, the corresponding original video images, video camera information, and other data into the database, and transmit them to the display module to facilitate users' real-time and accurate monitoring of natural resources.

[0079] This invention also provides a surveillance video image event detection system combined with a heterogeneous electronic fence, comprising the following modules:

[0080] The image acquisition module is used to acquire real-time monitoring video images by configuring the camera cruise strategy and obtain cruise image information;

[0081] The land parcel calculation module is used to calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel.

[0082] The video electronic fence acquisition module is used to convert the land parcels that need to be monitored into video electronic fences based on the information from the patrol images and the internal and external parameters of the camera.

[0083] The event detection module is used to perform event detection on cruise images using the Mask R-CNN network. The specific processing procedure of the Mask R-CNN network is as follows: the input image extracts a shared feature map through a backbone network composed of a ResNet residual network and a Feature Pyramid Network (FPN). This feature map is then input into the Region Proposal Network (RPN) to obtain candidate bounding box regions. The candidate bounding box regions are filtered based on video geofencing. The filtered regions are then input into the Region of Interest (ROI) matching layer (RoIAlign). RoIAlign uses bilinear interpolation to replace the quantization operation in the feature aggregation process. Finally, the ROI is input into the mask branch and the FAST-RCNN branch respectively. The final network outputs the predicted mask bounding box vertex coordinates, the predicted target bounding box center point coordinates and length and width data, and the predicted target bounding box category.

[0084] The monitoring module is used to merge and store the event results detected by video images, the corresponding original video images, and video camera information data into the database, and then transmit them to the display module to facilitate users' real-time and accurate monitoring of natural resources.

[0085] The specific implementation methods of each module are the same as those of each step, and will not be described in this invention.

[0086] This invention presents an event detection method combining heterogeneous electronic fences and surveillance video images. The invention dynamically transforms geographic electronic fences into video electronic fences in real time using geographic video coordinate transformation technology. Simultaneously, it combines this with real-time acquisition of patrol images from surveillance video cameras and utilizes the Mask R-CNN algorithm to perform event detection on these patrol images, thereby achieving real-time and accurate monitoring of natural resources.

[0087] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the invention.

Claims

1. A method for detecting events in surveillance video images combined with heterogeneous electronic fences, characterized in that, Includes the following steps: Step 1: By configuring the camera patrol strategy, real-time monitoring video images are collected to obtain patrol image information; Step 2: Calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel; Step 3: Based on the information from the patrol images and the internal and external parameters of the camera, convert the land parcels that need to be monitored into video electronic fences; Step 4: Use the Mask R-CNN network to perform event detection on the cruise images. The specific processing procedure of the Mask R-CNN network is as follows: the input image extracts a shared feature map through the backbone network composed of the Residual Network (ResNet) and the Feature Pyramid Network (FPN). This feature map is then input into the Region Proposal Network (RPN) to obtain candidate bounding box regions. The candidate bounding box regions are filtered based on the video geofence in Step 3. The filtered regions are then input into the Region of Interest (ROI) matching layer (RoIAlign). RoIAlign uses bilinear interpolation to replace the quantization operation in the feature aggregation process. Finally, the ROI is input into the mask branch and the FAST-RCNN branch respectively. The final network outputs the predicted mask bounding box vertex coordinates, the predicted target bounding box center point coordinates and length and width data, and the predicted target bounding box category.

2. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: The cruise image information in step 1 is V ptz ,in: p∈[0,360] t∈[-180,180] z∈[1,50] p represents the range of horizontal movement of the camera; t represents the range of vertical movement of the camera; z represents the zoom range.

3. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: The land parcels in step 2 are represented as follows: O m =MultiPolygon m (((Lng n Lat n ))) m∈[1,M] n∈[1,N] Where M represents the number of monitored land parcels within the camera's field of view, and N represents the number of polygon sides corresponding to the monitored land parcels; Lng n and Lat n MultiPolygon represents the latitude and longitude coordinates of the vertices of a polygon. m (((Lng n Lat n ))) is a way of representing polygons.

4. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: Step 3 first constructs a mapping model between geographic coordinates and video images under camera device posture changes. Then, combining the camera parameters captured by the image, the mapping model is used to convert the land parcels into video fences. The video electronic fence is represented as follows: O m =MultiPolygon m (((u n v n ))) m∈[1,M] n∈[1,N] Among them, u n and v n This represents the pixel coordinates of the video image under the pan-tilt-z (PTZ) mode of the camera corresponding to the polygon, where p represents the range of horizontal movement of the camera; t represents the range of vertical movement of the camera; z represents the zoom range; M represents the number of monitored land parcels within the camera's field of view; and N represents the number of polygon sides corresponding to the monitored land parcels. MultiPolygon m (((Lng n Lat n ))) is a way of representing polygons.

5. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 4, characterized in that: In the actual conversion process, u is filtered out. n and v n The occurrence of negative values.

6. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: The input image in step 4 is the cruise image obtained in step 1; after obtaining the candidate box region, the region with high score is calculated using the non-maximum suppression (NMS) algorithm, and regions with high overlap, exceeding the image boundary, and low confidence are removed.

7. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: In step 4, the video electronic fence is proportionally converted into a fence with the corresponding feature size. Then, the intersection of the fence with the candidate box area with the corresponding feature size is calculated. The area with an intersection area greater than a certain threshold is retained to complete the area filtering.

8. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: The RoIAlign processing flow is as follows: First, iterate through the selected candidate regions while keeping the floating-point numbers unchanged, divide the candidate regions into k*k units, calculate four fixed coordinate values ​​in each unit, calculate the values ​​of these four positions using bilinear interpolation, cancel the quantization process, and use the max pooling method to summarize the position values ​​to obtain the region of interest.

9. The method for detecting events in surveillance video images combined with heterogeneous electronic fences as described in claim 1, characterized in that: The mask branch includes deconvolutional and convolutional layers; the FAST-RCNN branch has fully connected layers, as well as bounding box classification and location regression operations, using the following loss function L: L=L cls +L box +L mask Among them, the border segmentation error L mask This refers to the error in the segmentation mask, and the error in the border position L. box Calculate the positional error and classification error L of the predicted bounding box output by the Mask R-CNN network. cls Calculate the error in the predicted object category output of the Mask R-CNN network.

10. A surveillance video image event detection system combined with a heterogeneous electronic fence, characterized in that, Includes the following modules: The image acquisition module is used to acquire real-time monitoring video images by configuring the camera cruise strategy and obtain cruise image information; The land parcel calculation module is used to calculate the land parcel monitored by each camera based on the positional relationship between the camera and the monitored land parcel. The video electronic fence acquisition module is used to convert the land parcels that need to be monitored into video electronic fences based on the information from the patrol images and the internal and external parameters of the camera. The event detection module is used to perform event detection on cruise images using the Mask R-CNN network. The specific processing procedure of the Mask R-CNN network is as follows: the input image extracts a shared feature map through a backbone network composed of a ResNet residual network and a Feature Pyramid Network (FPN). This feature map is then input into the Region Proposal Network (RPN) to obtain candidate bounding box regions. The candidate bounding box regions are filtered based on video geofencing. The filtered regions are then input into the Region of Interest (ROI) matching layer (RoIAlign). RoIAlign uses bilinear interpolation to replace the quantization operation in the feature aggregation process. Finally, the ROI is input into the mask branch and the FAST-RCNN branch respectively. The final network outputs the predicted mask bounding box vertex coordinates, the predicted target bounding box center point coordinates and length and width data, and the predicted target bounding box category. The monitoring module is used to merge and store the event results detected by video images, the corresponding original video images, and video camera information data into the database, and then transmit them to the display module to facilitate users' real-time and accurate monitoring of natural resources.

Images