Cable anti-winding intelligent monitoring method and device

Abstract

The present application relates to the field of computer vision and deep learning, and specifically provides a cable anti-winding intelligent monitoring method and device, which has the following steps: first, a target detection network adopts a DCNv2 deformable convolution network to enhance a YOLOv5 backbone network, so as to improve the detection accuracy of the minimum bounding rectangle of the human target; the IoU intersection-over-union ratio of the human detection frame and the preset warning area is calculated to make a risk prediction: when the IoU is less than a threshold alpha, it is determined that the target has not entered the warning or winding area, and the subsequent inference process is terminated; when the IoU is greater than or equal to the threshold alpha, a collaborative inference mechanism is triggered; then a segmentation verification network is performed, a CBAM module is embedded into the MODNet architecture based on a composite attention mechanism, and fine segmentation of the human ROI area is realized. Compared with the prior art, the present application can solve the technical problem of insufficient human target detection and warning accuracy in the overhead monitoring scene.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning, specifically providing a method and device for intelligent monitoring of cable anti-winding enhanced by top-down human image detection and semantic segmentation. Background Technology

[0002] In overhead monitoring scenarios, due to the trapezoidal distortion effect caused by the perspective projection of the camera (the human torso exhibits a geometric distortion that is wider at the top and narrower at the bottom), traditional monitoring methods based on rectangular detection boxes have significant technical defects: First, the rectangular detection box is difficult to fit the distorted human body contour, resulting in a redundant background area accounting for more than 30%, which seriously affects the target positioning accuracy; Second, slight movements of human limbs can easily cause severe shaking of the detection box (the cross-union ratio fluctuates by more than ±15%), causing the warning judgment mechanism based on the cross-union ratio (IoU) to generate high-frequency false alarms.

[0003] Furthermore, in order to balance detection accuracy and speed, a single network model architecture is difficult to meet the high accuracy requirements in industrial scenarios, while a multi-network serial model architecture is difficult to meet the strict requirement of algorithm latency of less than 40ms. Summary of the Invention

[0004] This invention addresses the shortcomings of the prior art by providing a highly practical intelligent monitoring method for preventing cable entanglement.

[0005] A further technical objective of this invention is to provide a reasonably designed, safe, and applicable intelligent monitoring device for preventing cable entanglement.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] The intelligent monitoring method for preventing cable entanglement includes the following steps:

[0008] First, for the target detection network, the DCNv2 deformable convolutional network is used to enhance the YOLOv5 backbone network, thereby improving the detection accuracy of the minimum bounding rectangle of human targets.

[0009] Risk prediction is made by calculating the intersection-over-union (IoU) ratio between the human detection bounding box and the preset warning area.

[0010] When IoU < threshold α, it is determined that the target has not entered the warning or entanglement area, and the subsequent reasoning process is terminated; when IoU ≥ threshold α, the collaborative reasoning mechanism is triggered.

[0011] Then, a segmentation verification network is constructed, which embeds the compound attention mechanism CBAM module based on the MODNet architecture to achieve fine segmentation of the human ROI region.

[0012] Furthermore, in the object detection network, a dataset is collected, the original image resolution is set to 1920×1080, the network input is 640×640, and the dataset is divided into training set, validation set and test set in an 8:1:1 ratio;

[0013] During the data annotation process, foreign objects held by the person are removed from the image detection frame, which is an angled detection frame.

[0014] Furthermore, color dithering, horizontal flipping, and mosaic enhancement data augmentation methods are employed in the data channel to add DCNv2 to the YOLOv5 backbone network. The DCNv2 formula is as follows:

[0015]

[0016] Where x is the input feature map, K is the convolution kernel size, and w K Here, p represents the weights of the convolution kernel, and p is the output position of the standard convolution. k It is the preset rule sampling offset, Δp k The learnable spatial offset, Δm k Learnable modulation factor.

[0017] Furthermore, the acquired original image, after being resized proportionally to 640×640, is input into the detection model to obtain a human image warning bounding box. Risk prediction is then made by calculating the Intersection over Union (IoU) between the human detection bounding box and the preset warning area.

[0018] When IoU < threshold α, it is determined that the target has not entered the warning or entanglement area, and the subsequent inference process is terminated; when IoU ≥ threshold α, the collaborative inference mechanism is triggered, and the segmentation verification network is entered.

[0019] Furthermore, in the segmentation validation network, the collected dataset consists of images based on regions of interest from human detection boxes. The images have different resolutions, but if the batch size is greater than 1, the input images must have the same resolution. The dataset is divided into training, validation, and test sets in a ratio of 8:1:1.

[0020] Furthermore, during the data annotation process, foreign objects such as sticks, cardboard, and rags were excluded from the human face mask. Color dithering, cutout, or mosaic enhancement data augmentation methods were used in the data channel. The cutout operation randomly cropped out regions from the image, with the number of cropped regions being random from 1 to 10, and the length and width of the cropped regions being random from 1 to 24, thus destroying the image integrity to achieve a regularization effect. Mosaic enhancement was applied to the segmentation network, and mosaic stitching was performed on both the input image and the label image to maintain pixel-level matching.

[0021] Furthermore, the segmentation model backbone uses the lightweight MODNet model. Based on the pre-trained model, the network model is finely sculpted using the collected dataset and a small learning rate. At the same time, a composite attention mechanism is embedded into the MODNet model to improve the limb edge segmentation effect.

[0022] The composite attention mechanism includes a channel attention module and a spatial attention module;

[0023] The channel attention module is used to model the dependencies between channels and generate a channel attention graph.

[0024] The spatial attention module is used to model the dependencies between spatial locations and generate a spatial attention map. The final output feature is the result of the element-wise multiplication of the input feature and the two attention maps.

[0025] Furthermore, calculate the IoU between the current frame and the human detection bounding box of the previous frame. If the IoU is greater than or equal to the threshold β, then reuse the historical segmentation results.

[0026] When IoU < threshold β, a lightweight segmentation network is started for pixel-level verification. The region of interest of the human image is resized to 224×224 and then input into the segmentation network to obtain the segmentation result.

[0027] Furthermore, the area ratio γ of the human face mask inside and outside the warning area is calculated:

[0028] When γ ≥ threshold δ, it is determined to be a high-risk state and an alarm signal is triggered; when γ < threshold δ, the safe state is maintained.

[0029] The intelligent monitoring device for preventing cable entanglement includes: at least one memory and at least one processor;

[0030] The at least one memory is used to store a machine-readable program;

[0031] The at least one processor is used to call the machine-readable program to execute the intelligent monitoring method for preventing cable entanglement.

[0032] Compared with the prior art, the intelligent monitoring method and device for preventing cable entanglement of the present invention have the following outstanding advantages:

[0033] This invention solves the technical problem of insufficient accuracy in human target detection and early warning in top-down monitoring scenarios, while ensuring that the overall system latency is ≤40ms.

[0034] Introducing DCNv2 into the YOLOv5 backbone network allows the sampling positions of the convolutional kernels to be dynamically adjusted based on the image content, effectively improving the network's adaptability to large changes in portraits at downward angles.

[0035] The YOLO series of mosaic enhancement algorithms are applied to the segmentation validation network, simultaneously performing mosaic stitching on the input and label images to maintain pixel-level matching between the input and labels. This significantly enhances data diversity and improves segmentation accuracy.

[0036] By embedding a composite attention mechanism into the MODNet network, which combines channel attention module and spatial attention module, we focus on limb edge features, thereby improving the limb edge segmentation effect and achieving a final segmentation accuracy of 95.6%.

[0037] A dual-threshold risk decision-making mechanism is introduced in cable anti-winding scenarios. By cascading judgments of two levels of IoU thresholds (α,β) and mask proportion threshold δ, the decision-making accuracy is greatly improved.

[0038] First, a primary screening mechanism is used to filter out target detection boxes with an IoU exceeding a threshold in the preset warning region. Then, through inter-frame consistency verification, historical segmentation results are directly reused for images with high similarity to the previous frame to avoid redundant calculations. This strategy enables dynamic allocation of computing resources, reduces redundant computation, and ensures that the overall system latency is ≤40ms. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating an intelligent monitoring method for preventing cable entanglement.

[0041] Figure 2 This is a schematic diagram of a composite attention mechanism in an intelligent monitoring method for preventing cable entanglement. Detailed Implementation

[0042] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] The following is a preferred embodiment:

[0044] like Figure 1As shown, the intelligent monitoring method for preventing cable entanglement in this embodiment includes the following steps:

[0045] First, for the target detection network, the DCNv2 deformable convolutional network is used to enhance the YOLOv5 backbone network, thereby improving the detection accuracy of the minimum bounding rectangle of human targets.

[0046] Risk prediction is made by calculating the intersection-over-union (IoU) ratio between the human detection bounding box and the preset warning area.

[0047] When IoU < threshold α, it is determined that the target has not entered the warning or entanglement area, and the subsequent reasoning process is terminated; when IoU ≥ threshold α, the collaborative reasoning mechanism is triggered.

[0048] Then, a segmentation verification network is constructed, which embeds the composite attention mechanism CBAM module based on the MODNet architecture to achieve fine segmentation of the human ROI region.

[0049] In the object detection network, a total of 6000 images were collected. The original image resolution was 1920×1080, and the network input was 640×640. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. To address the influence of foreign objects such as sticks, cardboard, and rags, these objects were excluded from the image detection bounding box during data annotation. An angled detection box was used, which reduced background interference to some extent compared to traditional rectangular detection boxes. Data augmentation methods such as color dithering, horizontal flipping, and mosaic enhancement were employed in the data channels to enhance the diversity of the input data.

[0050] Due to the complex industrial environment and the downward viewing angle, the human figure is not fixed but constantly changing. To improve the network's adaptability to these changes, DCNv2 was added to the YOLOv5 backbone network.

[0051] The formula for DCNv2 is as follows:

[0052]

[0053] Where x is the input feature map, K is the convolution kernel size (e.g., K is 9 for 3×3), and w K Here, p represents the weights of the convolution kernel, and p is the output position of the standard convolution. k It is the preset rule sampling offset, Δp k The learnable spatial offset (deformation offset), Δm k Learnable modulation factors (weighting factors). DCNv2 makes the sampling position of the convolution kernel no longer fixed, but dynamically adjusted according to the image content, effectively improving the network's adaptability to changes.

[0054] The original images are resized to 640×640 and then input into the detection model to obtain human image warning bounding boxes. Risk prediction is then performed by calculating the IoU (Intersection over Union) between the human detection bounding box and the preset warning area.

[0055] When IoU < threshold α, it is determined that the target has not entered the warning / entanglement area, and the subsequent inference process is terminated; when IoU ≥ threshold α, the collaborative inference mechanism is triggered, and the segmentation verification network is entered.

[0056] In the segmentation validation network, a total of 2000 images were collected. All images were regions of interest based on human face detection boxes. The image resolutions varied, but if the batch size was greater than 1, the input images needed to have the same resolution (the resolution used in this invention is 224×224). The dataset was divided into training set, validation set and test set in a ratio of 8:1:1.

[0057] Similarly, to eliminate the influence of foreign objects, sticks, cardboard, rags, and other foreign objects are excluded from the portrait mask during data annotation. Data augmentation methods such as color dithering, Cutout, and mosaic enhancement are employed in the data channels. Cutout involves randomly cropping regions from the image, with the number of cropped regions ranging from 1 to 10, and the length and width of each cropped region ranging from 1 to 24, thus disrupting image integrity for regularization. Mosaic enhancement, a data augmentation method proposed by the YOLO series, is applied to the segmentation network in this invention. Mosaic stitching is performed simultaneously on the input and label images, maintaining pixel-level matching. The segmentation backbone uses the lightweight MODNet model, which focuses on portrait segmentation and achieves good results from a frontal view. To adapt to the overhead view in industrial scenarios, the network model is finely sculpted using the collected dataset and a small learning rate (<0.005) based on the pre-trained model. Furthermore, a composite attention mechanism is embedded into the MODNet model to improve limb edge segmentation.

[0058] like Figure 2 As shown, the composite attention mechanism consists of two parts: a ChannelAttention Module, which models the dependencies between channels and generates a channel attention map; and a SpatialAttention Module, which models the dependencies between spatial locations and generates a spatial attention map. The final output feature is the element-wise multiplication of the input feature with the two attention maps. Embedding the composite attention mechanism effectively improves the limb edge segmentation performance, achieving a final segmentation accuracy of 95.6%.

[0059] Calculate the IoU between the current frame and the human detection bounding box in the previous frame. If IoU ≥ threshold β, reuse the historical segmentation results to avoid redundant calculations; if IoU < threshold β, activate the lightweight segmentation network for pixel-level verification. Resize the region of interest (ROI) of the human image to 224×224 and input it into the segmentation network to obtain the segmentation result.

[0060] Calculate the area ratio γ of the human face mask inside and outside the warning area:

[0061] When γ ≥ threshold δ, it is determined to be a high-risk state (entering the warning / entanglement area) and an alarm signal is triggered; when γ < threshold δ, the safe state determination is maintained.

[0062] Based on the above method, the cable anti-winding intelligent monitoring device in this embodiment includes: at least one memory and at least one processor;

[0063] The at least one memory is used to store a machine-readable program;

[0064] The at least one processor is used to call the machine-readable program to execute the intelligent monitoring method for preventing cable entanglement.

[0065] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0066] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0067] The above-described specific embodiments are merely specific examples of the present invention. The patent protection scope of the present invention includes, but is not limited to, the above-described specific embodiments. Any technical solution that conforms to the above-described specific embodiments of the present invention and any appropriate changes or substitutions made by those skilled in the art should fall within the patent protection scope of the present invention.

[0068] Although embodiments of the invention have been shown and described, those skilled in the art will find them difficult to understand.

[0069] For those familiar with the invention, it is understood that modifications can be made without departing from the principles and spirit of the invention.

[0070] These embodiments may be subject to various changes, modifications, substitutions, and variations, and the scope of the invention is defined by the appended claims.

[0071] Claims and their equivalents are limited.

Claims

1. A smart monitoring method for preventing cable entanglement, characterized in that, It has the following steps: First, for the target detection network, the DCNv2 deformable convolutional network is used to enhance the YOLOv5 backbone network, thereby improving the detection accuracy of the minimum bounding rectangle of human targets. Risk prediction is made by calculating the intersection-over-union (IoU) ratio between the human detection bounding box and the preset warning area. When IoU < threshold α, it is determined that the target has not entered the warning or entanglement area, and the subsequent reasoning process is terminated; when IoU ≥ threshold α, the collaborative reasoning mechanism is triggered. Then, a segmentation verification network is constructed, which embeds the composite attention mechanism CBAM module based on the MODNet architecture to achieve fine segmentation of the human ROI region; The segmentation model backbone uses the lightweight MODNet model. Based on the pre-trained model, the network model is finely sculpted using the collected dataset and a small learning rate. At the same time, a composite attention mechanism is embedded into the MODNet model to improve the limb edge segmentation effect. The composite attention mechanism includes a channel attention module and a spatial attention module; The channel attention module is used to model the dependencies between channels and generate a channel attention graph. The spatial attention module is used to model the dependencies between spatial locations and generate a spatial attention graph. The final output feature is the result of the element-wise multiplication of the input feature with the two attention graphs. Calculate the IoU between the current frame and the human detection bounding box in the previous frame. If the IoU is greater than or equal to the threshold β, then reuse the historical segmentation results. When IoU < threshold β, a lightweight segmentation network is started for pixel-level verification. The region of interest of the human image is resized to 224×224 and then input into the segmentation network to obtain the segmentation result. Calculate the area ratio γ of the human face mask inside and outside the warning area: When γ ≥ threshold δ, it is determined to be a high-risk state and an alarm signal is triggered; when γ < threshold δ, the safe state is maintained.

2. The intelligent monitoring method for preventing cable entanglement according to claim 1, characterized in that, In the object detection network, a dataset is collected, the original image resolution is set to 1920×1080, the network input is 640×640, and the dataset is divided into training set, validation set and test set in an 8:1:1 ratio. During the data annotation process, foreign objects held by the person are removed from the image detection frame, which is an angled detection frame.

3. The intelligent monitoring method for preventing cable entanglement according to claim 2, characterized in that, Color dithering, horizontal flipping, and mosaic enhancement data augmentation methods are used in the data channel to add DCNv2 to the YOLOv5 backbone network. The DCNv2 formula is as follows: ； Where x is the input feature map, K is the kernel size, wK is the kernel weight, p is the standard convolution output position, pk is the preset regular sampling offset, Δpk is the learnable spatial offset, and Δmk is the learnable modulation factor.

4. The intelligent monitoring method for preventing cable entanglement according to claim 3, characterized in that, The original images are resized to 640×640 and then input into the detection model to obtain human warning bounding boxes. Risk prediction is then made by calculating the Intersection over Union (IoU) between the human detection bounding box and the preset warning area. When IoU < threshold α, it is determined that the target has not entered the warning or entanglement area, and the subsequent inference process is terminated; when IoU ≥ threshold α, the collaborative inference mechanism is triggered, and the segmentation verification network is entered.

5. The intelligent monitoring method for preventing cable entanglement according to claim 4, characterized in that, In the segmentation validation network, the dataset is collected. The images are all regions of interest based on human image detection boxes. The image resolutions are different, but if the batch is greater than 1, the input images need to have the same resolution. The dataset is divided into training set, validation set and test set in a ratio of 8:1:

1.

6. The intelligent monitoring method for preventing cable entanglement according to claim 5, characterized in that, During the data annotation process, foreign objects such as sticks, cardboard, and rags were excluded from the human face mask. In the data channel, color dithering, cutout, or mosaic enhancement data augmentation methods were used. The cutout operation randomly cropped out regions from the image, with the number of cropped regions being random from 1 to 10, and the length and width of the cropped regions being random from 1 to 24, thus destroying the image integrity to achieve a regularization effect. Mosaic enhancement was applied to the segmentation network, and mosaic stitching was performed on both the input image and the label image to maintain pixel-level matching.

7. A cable anti-winding intelligent monitoring device, characterized in that, include: At least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to invoke the machine-readable program to execute the method according to any one of claims 1 to 6.

Images