Intelligent safety evaluation system and method for industry and trade operation scene

Abstract

The invention discloses an intelligent safety evaluation system and method for an industry and trade operation scene, and relates to the technical field of artificial intelligence. A video image set is obtained by acquiring a video stream of a target area in real time and performing frame extraction; enhancing the target video image to obtain a target enhanced image, identifying the target enhanced image to obtain a label, adding the label to the target enhanced image to obtain a target label graph, segmenting the target label graph to obtain a feature extraction graph, and performing occlusion judgment on the feature extraction graph to obtain a judgment result; if the judgment result is no shielding, carrying out safety evaluation to obtain an evaluation result, and carrying out risk alarm; and if the judgment result is shielding, carrying out risk alarm. Multi-camera real-time video stream acquisition and frame extraction are carried out, the calculation amount is reduced, the calculation force overload is avoided, the accuracy of target model identification is improved by enhancing the image, so that the target label graph is accurately generated, subsequent feature extraction and occlusion judgment are facilitated, differential processing of occlusion and non-occlusion conditions is realized, and the accuracy of target model identification is improved. And misjudgment and missed judgment caused by shielding are avoided.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an intelligent safety assessment system and method for industrial and commercial operation scenarios. Background Technology

[0002] With the rapid development of the industrial and commercial sectors, safety management at production sites faces increasingly complex challenges. Traditional safety management methods mainly rely on manual inspections and post-event analysis, which suffer from low efficiency, poor real-time performance, and strong subjectivity. In recent years, with the rapid development of emerging technologies such as the Internet of Things, artificial intelligence, big data, and edge computing, intelligent safety assessment technologies for industrial and commercial operation scenarios have emerged.

[0003] Publication No. CN119204686A discloses a method, system, and equipment for assessing and warning safety risks to construction workers. The method involves acquiring attribute data and on-site behavior data of construction workers as input data, inputting this data into a safety risk assessment model to obtain multidimensional feature values ​​of the input variables, performing cluster analysis on these feature values ​​to obtain a safety risk classification for construction workers, and generating a safety risk warning indication based on this classification.

[0004] In existing technologies, the large amount of data acquired leads to excessive computation, and in industrial and commercial scenarios, poor image quality results in slow target model recognition speed and low accuracy. In the security assessment process, video occlusion can easily lead to misjudgment or missed judgment. Summary of the Invention

[0005] The purpose of this invention is to address the problems mentioned in the background art, such as excessive computational load due to acquired data, slow target model recognition speed and low accuracy due to poor image quality in industrial and commercial scenarios, and misjudgment or omission due to video occlusion during the safety assessment process. Therefore, this invention proposes an intelligent safety assessment system and method for industrial and commercial operation scenarios.

[0006] A first aspect of this invention provides an intelligent safety assessment method for industrial and commercial operation scenarios, the method comprising: The video stream of the target area is acquired in real time, and the video stream is extracted frame by frame to obtain a video image set; the video image set contains multiple video images. A target enhanced image is obtained by enhancing a target video image; a target model is used to identify labels in the target enhanced image; and the labels are added to the target enhanced image to obtain a target label map; the target video image can be any one of the video images in the set. The target label map is segmented according to the label to obtain a feature extraction map, and the feature extraction map is subjected to occlusion judgment to obtain a judgment result; the judgment result includes: occluded and unoccluded; If the assessment result is that the target is not obscured, a security assessment is performed on the target person to obtain the assessment result. If the assessment result is risky, a risk alert is issued. If the judgment result is obstruction, an audible and visual alarm will be triggered. If the obstruction is still determined within a preset time period, a risk alarm will be triggered.

[0007] Optionally, a label is obtained by identifying the target-enhanced image using a target model, wherein the target model is based on an improvement of YOLOv8 and includes: Replace the Conv modules in layers 0, 1, 3, 5 and 7 of the YOLOv8 model with CBS modules, and replace the neck structure of YOLOv8 with the target neck structure to obtain the target model; The workflow of the target neck structure includes: The output of the 9th layer SPPF module is used as the first input diagram of the target neck structure, the output of the 8th layer C2f module is used as the second input diagram of the target neck structure, the output of the 6th layer C2f module is used as the third input diagram of the target neck structure, the output of the 4th layer C2f module is used as the fourth input diagram of the target neck structure, and the output of the 5th layer C2f module is used as the fifth input diagram of the target neck structure. The first input image is input to the CBS module to obtain the first feature image. The fourth input image is input to the CBS module to obtain the second feature image. The first, second, and third input images are input to the adaptive fusion module to obtain the third feature image. The third feature image is sequentially input to the target C2F module and the CBS module to obtain the fourth feature image. The fifth input image is input to the CBS module to obtain the fifth feature image. The fourth, fourth, and fifth input images are input to the adaptive fusion module to obtain the sixth feature image. The 6th feature map is input into the C2f module to obtain the 7th feature map. The 2nd, 3rd, and 4th input maps are input into the adaptive fusion module to obtain the 8th feature map. The 7th and 8th feature maps are added together to obtain the 9th feature map. The 9th feature map is used as the input of the detection head. The 9th feature map is upsampled and then concatenated with the 5th input map to obtain the 10th feature map. The 10th feature map is then added to the 1st input map to obtain the 11th feature map. The 11th feature map is then used as the input of the detection head. The 6th feature map is input into the CBS module to obtain the 12th feature map. The 4th feature map and the 12th feature map are concatenated to obtain the 13th feature map. The 13th feature map is sequentially input into the target C2F module and the CBS module to obtain the 14th feature map. The 1st feature map and the 14th feature map are concatenated to obtain the 15th feature map. The 15th feature map is used as the input of the detection head.

[0008] Optional, the working principle of the adaptive fusion module includes: Obtain the original feature tensor, and downsample the original feature tensor to obtain high-scale feature tensor, medium-scale feature tensor and low-scale feature tensor; The high-scale feature tensor and the mesoscale feature tensor are input into the multi-scale fusion module to obtain the first tensor, and the mesoscale feature tensor and the low-scale feature tensor are input into the multi-scale fusion module to obtain the second tensor; The first tensor is input into the channel attention module to obtain the third tensor. The third tensor is concatenated with the second tensor to obtain the fourth tensor. The fourth tensor is concatenated with the low-scale feature tensor to obtain the fifth tensor. The fifth tensor is used as the output of the adaptive fusion module.

[0009] Optional improvements to the target C2f module include: The target C2f module is obtained by replacing the bottleneck module in the C2f module with the target bottleneck module. Improvements to the target bottleneck module include: The target bottleneck module is obtained by replacing the Conv module in the bottleneck module with the target Conv module. The working principle of the target Conv module includes: Obtain an initial task map, divide the initial task map into a first segmented map and a second segmented map, input the second segmented map into a 1×1 Conv module to obtain the first task map, input the first task map into a 3×3 Conv module to obtain the second task map, input the second task map into a 5×5 Conv module to obtain the third task map, and input the third task map into a 7×7 Conv module to obtain the fourth task map; The first segmentation image, the first operation image, the second operation image, the third operation image, and the fourth operation image are stitched together to obtain the fifth operation image. The fifth operation image is input into the 1×1Conv module to obtain the sixth operation image. The sixth operation image is used as the output of the target C2F module.

[0010] Optionally, the target label map is segmented according to the label to obtain a feature extraction map, including: A regional standard image is acquired, and the target label image is processed through a U-net network to obtain a task feature map. The task feature map is convolved to obtain a segmentation mask. The segmentation mask is subjected to a preset operation to obtain a target mask. The regional standard image is subjected to a preset operation to obtain a target standard image. The target mask and the target standard image are fused to obtain an illumination feature map. The preset operation includes convolution operation and upsampling operation. The illumination feature map and the job feature map are fused to obtain an illumination enhancement map, and the illumination enhancement map is segmented according to the label to obtain a feature extraction map.

[0011] A second aspect of this invention provides an intelligent safety assessment system for industrial and commercial operation scenarios, the system comprising: The image extraction module is used to acquire the video stream of the target area in real time, and extract frames from the video stream to obtain a video image set; the video image set contains multiple video images. An image recognition module is used to enhance a target video image to obtain a target enhanced image, recognize the target enhanced image using a target model to obtain a label, and add the label to the target enhanced image to obtain a target label map; the target video image is any one of the video images in the set; An occlusion determination module is used to segment the target label image according to the label to obtain a feature extraction image, and to determine the occlusion of the feature extraction image to obtain a determination result; the determination result includes: occluded and unoccluded; The first judgment module is used to perform a safety assessment on the target person if the judgment result is that the person is not obscured, and to issue a risk alert if the assessment result is that the person is at risk. The second judgment module is used to issue an audible and visual alarm if the judgment result is obstruction, and to issue a risk alarm if the obstruction is still determined within a preset time period.

[0012] Optionally, a label is obtained by identifying the target-enhanced image using a target model, wherein the target model is based on an improvement of YOLOv8 and includes: Replace the Conv modules in layers 0, 1, 3, 5 and 7 of the YOLOv8 model with CBS modules, and replace the neck structure of YOLOv8 with the target neck structure to obtain the target model; The workflow of the target neck structure includes: The output of the 9th layer SPPF module is used as the first input diagram of the target neck structure, the output of the 8th layer C2f module is used as the second input diagram of the target neck structure, the output of the 6th layer C2f module is used as the third input diagram of the target neck structure, the output of the 4th layer C2f module is used as the fourth input diagram of the target neck structure, and the output of the 5th layer C2f module is used as the fifth input diagram of the target neck structure. The first input image is input to the CBS module to obtain the first feature image. The fourth input image is input to the CBS module to obtain the second feature image. The first, second, and third input images are input to the adaptive fusion module to obtain the third feature image. The third feature image is sequentially input to the target C2F module and the CBS module to obtain the fourth feature image. The fifth input image is input to the CBS module to obtain the fifth feature image. The fourth, fourth, and fifth input images are input to the adaptive fusion module to obtain the sixth feature image. The 6th feature map is input into the C2f module to obtain the 7th feature map. The 2nd, 3rd, and 4th input maps are input into the adaptive fusion module to obtain the 8th feature map. The 7th and 8th feature maps are added together to obtain the 9th feature map. The 9th feature map is used as the input of the detection head. The 9th feature map is upsampled and then concatenated with the 5th input map to obtain the 10th feature map. The 10th feature map is then added to the 1st input map to obtain the 11th feature map. The 11th feature map is then used as the input of the detection head. The 6th feature map is input into the CBS module to obtain the 12th feature map. The 4th feature map and the 12th feature map are concatenated to obtain the 13th feature map. The 13th feature map is sequentially input into the target C2F module and the CBS module to obtain the 14th feature map. The 1st feature map and the 14th feature map are concatenated to obtain the 15th feature map. The 15th feature map is used as the input of the detection head.

[0013] Optional, the working principle of the adaptive fusion module includes: Obtain the original feature tensor, and downsample the original feature tensor to obtain high-scale feature tensor, medium-scale feature tensor and low-scale feature tensor; The high-scale feature tensor and the mesoscale feature tensor are input into the multi-scale fusion module to obtain the first tensor, and the mesoscale feature tensor and the low-scale feature tensor are input into the multi-scale fusion module to obtain the second tensor; The first tensor is input into the channel attention module to obtain the third tensor. The third tensor is concatenated with the second tensor to obtain the fourth tensor. The fourth tensor is concatenated with the low-scale feature tensor to obtain the fifth tensor. The fifth tensor is used as the output of the adaptive fusion module.

[0014] Optional improvements to the target C2f module include: The target C2f module is obtained by replacing the bottleneck module in the C2f module with the target bottleneck module. Improvements to the target bottleneck module include: The target bottleneck module is obtained by replacing the Conv module in the bottleneck module with the target Conv module. The working principle of the target Conv module includes: Obtain an initial task map, divide the initial task map into a first segmented map and a second segmented map, input the second segmented map into a 1×1 Conv module to obtain the first task map, input the first task map into a 3×3 Conv module to obtain the second task map, input the second task map into a 5×5 Conv module to obtain the third task map, and input the third task map into a 7×7 Conv module to obtain the fourth task map; The first segmentation image, the first operation image, the second operation image, the third operation image, and the fourth operation image are stitched together to obtain the fifth operation image. The fifth operation image is input into the 1×1Conv module to obtain the sixth operation image. The sixth operation image is used as the output of the target C2F module.

[0015] Optionally, the occlusion determination module includes: The feature extraction module is used to acquire a regional standard image, process the target label image through a U-net network to obtain a task feature map, convolve the task feature map to obtain a segmentation mask, perform a preset operation on the segmentation mask to obtain a target mask, perform a preset operation on the regional standard image to obtain a target standard image, and fuse the target mask and the target standard image to obtain an illumination feature map; the preset operation includes convolution operation and upsampling operation; The image segmentation module is used to fuse the illumination feature map and the job feature map to obtain an illumination enhancement map, and to segment the illumination enhancement map according to the label to obtain a feature extraction map.

[0016] The beneficial effects of this invention are: This invention proposes an intelligent safety assessment method for industrial and commercial operation scenarios. It reduces the amount of computation and avoids computing power overload by acquiring video streams in real time through multiple cameras and extracting frames. By enhancing the image, it improves the accuracy of target model recognition, thereby accurately generating target label maps, which is beneficial for subsequent feature extraction and occlusion judgment. It realizes differentiated processing of occluded and unoccluded situations and avoids misjudgment and missed judgment caused by occlusion. Attached Figure Description

[0017] Figure 1 A flowchart of an intelligent safety assessment method for industrial and commercial operation scenarios is provided as an embodiment of the present invention; Figure 2 This invention provides a framework diagram of an intelligent safety assessment system for industrial and commercial operation scenarios. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides an intelligent safety assessment method for industrial and commercial operation scenarios. See also... Figure 1 , Figure 1 A flowchart illustrating an intelligent safety assessment method for industrial and commercial operation scenarios provided by an embodiment of the present invention. The method includes the following steps: S101, real-time acquisition of video stream of target area, and extraction of video stream frames to obtain video image set; S102, enhance the target video image to obtain the target enhanced image, identify the target enhanced image through the target model to obtain the label, and add the label to the target enhanced image to obtain the target label map; S103, the target label map is segmented according to the label to obtain the feature extraction map, and the occlusion judgment is performed on the feature extraction map to obtain the judgment result; S104. If the judgment result is that the target is not obscured, a security assessment is performed on the target person to obtain the assessment result. If the assessment result is risky, a risk alert is issued. S105 If the judgment result is obstruction, then an audible and visual alarm will be triggered; if the obstruction is still determined within a preset time period, then a risk alarm will be triggered.

[0020] The video image set contains multiple video images; the target video image is any one of the video image sets; the judgment result includes: occluded and unoccluded; The intelligent safety assessment method for industrial and commercial operation scenarios provided by this invention uses multiple cameras to collect video streams in real time and extract frames, reducing the amount of computation and avoiding computational overload. By enhancing the image, the accuracy of target model recognition is improved, thereby accurately generating target label maps, which is beneficial for subsequent feature extraction and occlusion judgment. It realizes differentiated processing of occluded and unoccluded situations, avoiding misjudgment and missed judgment caused by occlusion.

[0021] In one implementation, the industrial and commercial operation scenario consists of multiple cameras. These cameras are responsible for collecting data on the work of personnel within the area and identifying whether their actions are compliant. Intelligent safety assessment of industrial and commercial operation scenarios refers to the management process of collecting data in real time on production operation sites in industrial and commercial industries (such as manufacturing, metallurgy, building materials, textiles, and machinery processing) through intelligent technologies such as the Internet of Things, AI, big data, and edge computing, thereby achieving automatic risk identification, hierarchical early warning of hidden dangers, and dynamic assessment of safety status.

[0022] In one implementation, video streams of the target area (such as workshop production lines or warehouse operation areas) are acquired in real time, and the video streams are processed by frame extraction to generate a set of video images. After multiple cameras cover the operation area, the real-time video streams can completely record the personnel's work trajectory (such as whether they are close to dangerous equipment or whether they operate tools according to regulations). The frame extraction operation can ensure the timeliness of data while reasonably controlling the amount of computation per frame image, avoiding computational overload caused by processing too many video streams, and ensuring that each frame image can be transmitted to subsequent modules in a timely manner, so as not to miss the critical moments of the operator's violation of regulations (such as temporarily crossing safety railings or touching materials without wearing protective gloves).

[0023] In one implementation, any target video image in the video image set is enhanced to generate a target enhanced image, and then the target model is used to identify the label and generate a target label map. This solves the recognition problem caused by poor image quality in industrial and commercial scenarios. Industrial and commercial operation sites often have problems such as uneven lighting (e.g., backlit corners of workshops, strong light interference from welding operations) and blurry images (e.g., dust contamination of camera lenses). Image enhancement (e.g., contrast adjustment, noise reduction) can repair the quality defects of the original image, making personnel outlines, protective equipment details, and operation actions clearer. The target model can more accurately identify the label and bind the label with the enhanced image to generate a target label map.

[0024] In one implementation, a safety assessment of the target person is conducted to obtain the assessment result. This is achieved by training an operation detection model using a Long Short-Term Memory (LSTM) network model, which is then used to assess the safety of the target person's work. The assessment result includes risk and normal. The model is trained using a dataset of operation image sequences labeled with compliance and non-compliance (such as not wearing protective equipment or touching equipment in violation of regulations). The parameters are optimized to minimize the classification loss. New operation image sequences are processed in the same way and then input into the trained LTM network model. The model outputs the operation category probability. If the non-compliance probability exceeds a preset threshold, the operation is judged as a violation, thus achieving image-based operation compliance detection. Occlusion detection is used to detect whether the operator's operation is unrecognizable. For example, if the operator always has their back to the camera, the operation cannot be detected.

[0025] In one implementation, after obtaining the target label map, a feature extraction map is generated based on the label segmentation, and the feature extraction map is subjected to occlusion judgment. Then, differentiated processing is performed for the two results of occlusion and non-occlusion, namely, audible and visual alarm (when occluded) and safety assessment (when not occluded). Since occlusion can easily lead to misjudgment and missed judgment, a risk alarm is triggered if the operator cannot be identified within a preset time period. For example, if the operator is not identified within 1 minute, a notification is sent through the workshop audible and visual alarm and the manager's mobile phone.

[0026] In one embodiment, a label is obtained by identifying the target-enhanced image using a target model. The target model is based on an improvement of YOLOv8 and includes: Replace the Conv modules in layers 0, 1, 3, 5 and 7 of the YOLOv8 model with CBS modules, and replace the neck structure of YOLOv8 with the target neck structure to obtain the target model; The workflow for the target neck structure includes: The output of the 9th layer SPPF module is used as the first input diagram of the target neck structure, the output of the 8th layer C2f module is used as the second input diagram of the target neck structure, the output of the 6th layer C2f module is used as the third input diagram of the target neck structure, the output of the 4th layer C2f module is used as the fourth input diagram of the target neck structure, and the output of the 5th layer C2f module is used as the fifth input diagram of the target neck structure. The first input image is input into the CBS module to obtain the first feature image. The fourth input image is input into the CBS module to obtain the second feature image. The first, second, and third input images are input into the adaptive fusion module to obtain the third feature image. The third feature image is sequentially input into the target C2F module and the CBS module to obtain the fourth feature image. The fifth input image is input into the CBS module to obtain the fifth feature image. The fourth, fourth, and fifth input images are input into the adaptive fusion module to obtain the sixth feature image. The 6th feature map is input into the C2f module to obtain the 7th feature map. The 2nd, 3rd, and 4th input maps are input into the adaptive fusion module to obtain the 8th feature map. The 7th and 8th feature maps are added together to obtain the 9th feature map. The 9th feature map is used as the input to the detection head. The 9th feature map is upsampled and then concatenated with the 5th input map to obtain the 10th feature map. The 10th feature map is then added to the 1st input map to obtain the 11th feature map. The 11th feature map is then used as the input of the detection head. The 6th feature map is input into the CBS module to obtain the 12th feature map. The 4th and 12th feature maps are concatenated to obtain the 13th feature map. The 13th feature map is input into the target C2F module and the CBS module in sequence to obtain the 14th feature map. The 1st and 14th feature maps are concatenated to obtain the 15th feature map. The 15th feature map is used as the input of the detection head.

[0027] In one implementation, the feature extraction and fusion capabilities of the model are improved by replacing the Conv module in the YOLOv8 model with the CBS module and introducing an adaptive fusion module. The CBS module can process feature maps more effectively, while the adaptive fusion module can flexibly integrate feature information from different levels, enabling the model to better capture multi-scale target features and improve the detection accuracy of targets of different sizes. The CBS module consists of convolutional layers (used to extract features, learning spatial features through sliding windows and parameter sharing), batch normalization layers (normalizing the output of convolutional layers to accelerate the training process and improve model stability), and activation layers (usually using Leaky ReLU or SiLU (Swish), etc., to introduce nonlinearity and enhance the expressive power of the model).

[0028] In one implementation, the introduction of the target neck structure optimizes the hierarchical structure of the model, enabling feature information to flow and fuse more efficiently within the model. By rationally arranging the processing order of the input map and feature map, the model maintains high performance while reducing redundant computation, improving running efficiency and detection performance, and making the model more stable and reliable when dealing with complex scenarios.

[0029] In one implementation, by upsampling and concatenating the feature maps, the model can better utilize the detailed information in the shallow feature maps, thereby enhancing the detection capability of small targets. When processing small targets, it can obtain richer feature representations, improving the accuracy and robustness of small target detection.

[0030] In one embodiment, the working principle of the adaptive fusion module includes: Obtain the original feature tensor, and then downsample the original feature tensor to obtain the high-scale feature tensor, the medium-scale feature tensor, and the low-scale feature tensor; The high-scale feature tensor and the mesoscale feature tensor are input into the multi-scale fusion module to obtain the first tensor, and the mesoscale feature tensor and the low-scale feature tensor are input into the multi-scale fusion module to obtain the second tensor; The first tensor is input into the channel attention module to obtain the third tensor. The third tensor is concatenated with the second tensor to obtain the fourth tensor. The fourth tensor is concatenated with the low-scale feature tensor to obtain the fifth tensor. The fifth tensor is used as the output of the adaptive fusion module.

[0031] In one implementation, high- and medium-scale features are fused, and then medium- and low-scale features are fused to ensure the coherence of feature fusion and gradient flow. This allows deep, high-level semantic information to be transmitted and enhances medium- and shallow features. The module integrates information from different depths and scales through a splicing operation, so that the output feature map contains both rich semantic information (which is beneficial for identifying large targets) and high-resolution detail information (which is beneficial for locating and identifying small targets), thereby improving the overall detection performance of the model for targets with huge size differences.

[0032] In one implementation, at the end of the fusion process, the fourth tensor containing intermediate semantics is concatenated with the original low-scale feature tensor. This directly introduces the network's lowest-level, highest-resolution, and most detailed features (such as edges and textures) into the features to be used for prediction. This compensates for the loss of small target information and low-level details that naturally occurs when the network becomes deeper, and enhances the ability to detect small targets, such as the hand movements of workers.

[0033] In one implementation, a multi-scale fusion module performs independent convolution and batch normalization on feature tensors from the network backbone at high, medium, and low scales. It then uses upsampling or downsampling operations to unify these tensors to the same spatial size, and finally integrates them into a comprehensive feature representation through concatenation. This allows the multi-scale fusion module to combine shallow features containing detailed information, mid-level features containing intermediate semantics, and deep features containing global contextual information, thereby enhancing the model's ability to capture features of targets of different sizes, including small targets. A channel attention module includes a channel attention branch and a position attention branch. The channel attention branch evaluates the importance of each feature channel through global average pooling and convolution, generating channel weights. The position attention branch performs pooling along the width and height directions respectively, and combines it with convolution to evaluate the importance of each spatial location, generating spatial weights. These two weights are fused through element-wise multiplication and weighted onto the original features, enabling the model to focus on the most information-rich feature channels and the most critical object spatial regions, improving the accuracy of target recognition and localization in small targets and complex scenes.

[0034] In one embodiment, the improvement to the target C2f module includes: The target C2f module is obtained by replacing the bottleneck module in the C2f module with the target bottleneck module. Improvements to the target bottleneck module include: The target bottleneck module is obtained by replacing the Conv module in the bottleneck module with the target Conv module. The working principle of the target Conv module includes: Obtain the initial task map, divide the initial task map into the first and second division maps, input the second division map into the 1×1 Conv module to obtain the first task map, input the first task map into the 3×3 Conv module to obtain the second task map, input the second task map into the 5×5 Conv module to obtain the third task map, and input the third task map into the 7×7 Conv module to obtain the fourth task map; The first segmentation image, the first task image, the second task image, the third task image, and the fourth task image are stitched together to obtain the fifth task image. The fifth task image is then input into the 1×1Conv module to obtain the sixth task image. The sixth task image is then used as the output of the target C2F module.

[0035] In one implementation, by introducing convolution kernels of different sizes (1×1, 3×3, 5×5, 7×7) to convolve different parts of the input feature map, it is possible to capture multi-scale spatial features from local to global, making the model more flexible in dealing with targets of different sizes and shapes, better adapting to the diversity of target sizes in target detection tasks, and improving the accuracy and robustness of detection.

[0036] In one implementation, the input channels are divided into multiple groups, and each group performs convolution operations independently. This reduces the number of channels that each convolution kernel needs to process. Although more convolutional layers are introduced, each convolution kernel has fewer parameters, which effectively reduces the number of model parameters, memory usage, and computational complexity. This makes the model more efficient during training and inference, and can improve the model's running speed and resource utilization without reducing performance.

[0037] In one implementation, the C2f module introduces a target Conv module and a target bottleneck module (an improvement based on the target Conv module), which enhances the multi-scale feature extraction capability. This effectively captures feature information from local to global perspectives, improving the model's detection accuracy and robustness for targets of different sizes.

[0038] In one embodiment, segmenting the target label map according to the label to obtain a feature extraction map includes: The process involves acquiring a regional standard image, processing the target label image using the U-net network to obtain a task feature map, convolving the task feature map to obtain a segmentation mask, performing preset operations on the segmentation mask to obtain a target mask, performing preset operations on the regional standard image to obtain a target standard image, and fusing the target mask and the target standard image to obtain an illumination feature map. The preset operations include convolution and upsampling operations. The illumination feature map and the task feature map are fused to obtain the illumination enhancement map. The illumination enhancement map is then segmented according to the labels to obtain the feature extraction map.

[0039] In one implementation, the regional standard image is a clear image obtained by controlling shooting conditions (such as adjusting light and optimizing exposure) under the same shooting scene and spatial layout as the image to be segmented. For example, in a surveillance scene, an image taken during the day when there is sufficient light can be used as the regional standard image for human segmentation at night or in a dim environment.

[0040] In one implementation, the target label map is processed by a U-net network to generate a task feature map. The encoder of the U-net network performs downsampling on the target label map to gradually extract high-level semantic features of the person. At the same time, the decoder performs upsampling and uses skip connections to fuse features from different levels of the encoder with corresponding levels of the decoder, finally outputting a task feature map containing fine details of the person. A convolution operation is performed on the acquired task feature map to generate a segmentation mask. A 1×1 convolution kernel is used to perform channel compression and feature integration on the task feature map, mapping the multi-channel task feature map into a single-channel binary segmentation mask (where 1 in the mask represents the predicted person region and 0 represents the background region), completing the initial separation of the person region from the background.

[0041] In one implementation, a preset operation of "convolution and upsampling" is performed on the initially generated segmentation mask to obtain the target mask. First, the segmentation mask is refined by a 3×3 convolution kernel to smooth the jagged edges of the mask. Then, upsampling is performed by bilinear interpolation or transposed convolution to adjust the size of the segmentation mask to be consistent with the standard image of the region. The target mask is a region filter that retains the illumination information corresponding to the region of mask 1 in the target standard image and removes the illumination data of the background region. Finally, an illumination feature map containing only the accurate illumination features of the human region is obtained.

[0042] In one implementation, the lighting feature map and the task feature map are fused to generate a lighting enhancement map. The lighting information of the character area in the lighting feature map is fused with the structural features of the character in the task feature map by using a weighted summation method to obtain a lighting enhancement map with uniform lighting and clear details.

[0043] Based on the same inventive concept, this invention also provides an intelligent safety assessment system for industrial and commercial operation scenarios. See also Figure 2 , Figure 2 A framework diagram of an intelligent safety assessment system for industrial and commercial operation scenarios provided in this embodiment of the invention includes: The image extraction module is used to acquire the video stream of the target area in real time, extract frames from the video stream to obtain a video image set; the video image set contains multiple video images. The image recognition module is used to enhance the target video image to obtain the target enhanced image, recognize the target enhanced image through the target model to obtain the label, and add the label to the target enhanced image to obtain the target label map; the target video image can be any one of the video images in the set; The occlusion detection module is used to segment the target label map according to the label to obtain the feature extraction map, and to perform occlusion detection on the feature extraction map to obtain the detection result; the detection result includes: occluded and unoccluded; The first judgment module is used to perform a safety assessment on the target person if the judgment result is that the person is not obscured, and to issue a risk alert if the assessment result is that the person is at risk. The second judgment module is used to issue an audible and visual alarm if the judgment result is obstruction, and to issue a risk alarm if the obstruction is still determined within a preset time period.

[0044] The intelligent safety assessment system for industrial and commercial operation scenarios provided by this invention uses multiple cameras to collect video streams in real time and extract frames, reducing the amount of computation and avoiding computational overload. By enhancing the image, the accuracy of target model recognition is improved, thereby accurately generating target label maps, which is beneficial for subsequent feature extraction and occlusion judgment. It realizes differentiated processing of occluded and unoccluded situations, avoiding misjudgments and omissions caused by occlusion.

[0045] In one embodiment, a label is obtained by identifying the target-enhanced image using a target model. The target model is based on an improvement of YOLOv8 and includes: Replace the Conv modules in layers 0, 1, 3, 5 and 7 of the YOLOv8 model with CBS modules, and replace the neck structure of YOLOv8 with the target neck structure to obtain the target model; The workflow for the target neck structure includes: The output of the 9th layer SPPF module is used as the first input diagram of the target neck structure, the output of the 8th layer C2f module is used as the second input diagram of the target neck structure, the output of the 6th layer C2f module is used as the third input diagram of the target neck structure, the output of the 4th layer C2f module is used as the fourth input diagram of the target neck structure, and the output of the 5th layer C2f module is used as the fifth input diagram of the target neck structure. The first input image is input into the CBS module to obtain the first feature image. The fourth input image is input into the CBS module to obtain the second feature image. The first, second, and third input images are input into the adaptive fusion module to obtain the third feature image. The third feature image is sequentially input into the target C2F module and the CBS module to obtain the fourth feature image. The fifth input image is input into the CBS module to obtain the fifth feature image. The fourth, fourth, and fifth input images are input into the adaptive fusion module to obtain the sixth feature image. The 6th feature map is input into the C2f module to obtain the 7th feature map. The 2nd, 3rd, and 4th input maps are input into the adaptive fusion module to obtain the 8th feature map. The 7th and 8th feature maps are added together to obtain the 9th feature map. The 9th feature map is used as the input to the detection head. The 9th feature map is upsampled and then concatenated with the 5th input map to obtain the 10th feature map. The 10th feature map is then added to the 1st input map to obtain the 11th feature map. The 11th feature map is then used as the input of the detection head. The 6th feature map is input into the CBS module to obtain the 12th feature map. The 4th and 12th feature maps are concatenated to obtain the 13th feature map. The 13th feature map is input into the target C2F module and the CBS module in sequence to obtain the 14th feature map. The 1st and 14th feature maps are concatenated to obtain the 15th feature map. The 15th feature map is used as the input of the detection head.

[0046] In one embodiment, the working principle of the adaptive fusion module includes: Obtain the original feature tensor, and then downsample the original feature tensor to obtain the high-scale feature tensor, the medium-scale feature tensor, and the low-scale feature tensor; The high-scale feature tensor and the mesoscale feature tensor are input into the multi-scale fusion module to obtain the first tensor, and the mesoscale feature tensor and the low-scale feature tensor are input into the multi-scale fusion module to obtain the second tensor; The first tensor is input into the channel attention module to obtain the third tensor. The third tensor is concatenated with the second tensor to obtain the fourth tensor. The fourth tensor is concatenated with the low-scale feature tensor to obtain the fifth tensor. The fifth tensor is used as the output of the adaptive fusion module.

[0047] In one embodiment, the improvement to the target C2f module includes: The target C2f module is obtained by replacing the bottleneck module in the C2f module with the target bottleneck module. Improvements to the target bottleneck module include: The target bottleneck module is obtained by replacing the Conv module in the bottleneck module with the target Conv module. The working principle of the target Conv module includes: Obtain the initial task map, divide the initial task map into the first and second division maps, input the second division map into the 1×1 Conv module to obtain the first task map, input the first task map into the 3×3 Conv module to obtain the second task map, input the second task map into the 5×5 Conv module to obtain the third task map, and input the third task map into the 7×7 Conv module to obtain the fourth task map; The first segmentation image, the first task image, the second task image, the third task image, and the fourth task image are stitched together to obtain the fifth task image. The fifth task image is then input into the 1×1Conv module to obtain the sixth task image. The sixth task image is then used as the output of the target C2F module.

[0048] In one embodiment, the occlusion determination module includes: The feature extraction module is used to acquire the regional standard image, process the target label image through the U-net network to obtain the task feature map, convolve the task feature map to obtain the segmentation mask, perform preset operations on the segmentation mask to obtain the target mask, perform preset operations on the regional standard image to obtain the target standard image, and fuse the target mask and the target standard image to obtain the illumination feature map; the preset operations include convolution operation and upsampling operation; The image segmentation module is used to fuse the illumination feature map and the task feature map to obtain the illumination enhancement map, and to segment the illumination enhancement map according to the label to obtain the feature extraction map.

[0049] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent safety assessment of a work trade operation scene, characterized in that, The method comprises: real-time acquisition of a video stream of a target area, frame extraction of the video stream to obtain a video image set; the video image set comprises a plurality of video images; enhancing a target video image to obtain a target enhanced image, identifying the target enhanced image through a target model to obtain a label, adding the label to the target enhanced image to obtain a target labeled image; the target video image is any one of the video image set; segmenting the target labeled image according to the label to obtain a feature extraction image, and performing occlusion judgment on the feature extraction image to obtain a judgment result; the judgment result includes: occlusion and no occlusion; if the judgment result is no occlusion, performing safety assessment on the target person to obtain an assessment result, and if the assessment result is risk, performing risk warning; if the judgment result is occlusion, performing sound and light warning, and if it is still determined to be occlusion within a preset time period, performing risk warning. 2.The intelligent safety assessment method of a work scene according to claim 1, wherein, The target model is based on the improvement of YOLOv8 and comprises: replacing the Conv module in the 0th layer, the 1st layer, the 3rd layer, the 5th layer and the 7th layer in the YOLOv8 model with a CBS module, and replacing the neck structure of YOLOv8 with a target neck structure to obtain a target model; the working process of the target neck structure comprises: taking the output of the 9th layer SPPF module as the 1st input image of the target neck structure, taking the output of the 8th layer C2f module as the 2nd input image of the target neck structure, taking the output of the 6th layer C2f module as the 3rd input image of the target neck structure, taking the output of the 4th layer C2f module as the 4th input image of the target neck structure, and taking the output of the 5th layer C2f module as the 5th input image of the target neck structure; inputting the 1st input image into the CBS module to obtain the 1st feature image, inputting the 4th input image into the CBS module to obtain the 2nd feature image, inputting the 1st feature image, the 2nd feature image and the 3rd input image into the adaptive fusion module to obtain the 3rd feature image, inputting the 3rd feature image into the target C2F module and the CBS module in sequence to obtain the 4th feature image, inputting the 5th input image into the CBS module to obtain the 5th feature image, and inputting the 4th input image, the 4th feature image and the 5th feature image into the adaptive fusion module to obtain the 6th feature image; inputting the 6th feature image into the C2f module to obtain the 7th feature image, inputting the 2nd input image, the 3rd input image and the 4th input image into the adaptive fusion module to obtain the 8th feature image, adding the 7th feature image and the 8th feature image to obtain the 9th feature image, and taking the 9th feature image as the input of the detection head; performing upsampling on the 9th feature image and splicing the 5th input image to obtain the 10th feature image, adding the 10th feature image and the 1st input image to obtain the 11th feature image, and taking the 11th feature image as the input of the detection head; The 6th feature map is input into a CBS module to obtain a 12th feature map, the 4th feature map and the 12th feature map are spliced to obtain a 13th feature map, the 13th feature map is sequentially input into a target C2F module and a CBS module to obtain a 14th feature map, the 1st feature map and the 14th feature map are spliced to obtain a 15th feature map, and the 15th feature map is taken as the input of the detection head. 3.The method of claim 1, wherein, The working principle of the adaptive fusion module includes: An original feature tensor is obtained, and the original feature tensor is down-sampled to obtain a high-scale feature tensor, a medium-scale feature tensor and a low-scale feature tensor; The high-scale feature tensor and the medium-scale feature tensor are input into a multi-scale fusion module to obtain a 1st tensor, and the medium-scale feature tensor and the low-scale feature tensor are input into a multi-scale fusion module to obtain a 2nd tensor; The 1st tensor is input into a channel attention module to obtain a 3rd tensor, the 3rd tensor and the 2nd tensor are spliced to obtain a 4th tensor, the 4th tensor and the low-scale feature tensor are spliced to obtain a 5th tensor, and the 5th tensor is taken as the output of the adaptive fusion module. 4.The method of claim 1, wherein The improvement of the target C2F module includes: The bottleneck module in the C2F module is replaced by a target bottleneck module to obtain a target C2F module. The improvement of the target bottleneck module includes: The Conv module in the bottleneck module is replaced by a target Conv module to obtain a target bottleneck module: The working principle of the target Conv module includes: An initial work graph is obtained, the initial work graph is segmented to obtain a 1st segmented graph and a 2nd segmented graph, the 2nd segmented graph is input into a 1×1 Conv module to obtain a 1st work graph, the 1st work graph is input into a 3×3 Conv module to obtain a 2nd work graph, the 2nd work graph is input into a 5×5 Conv module to obtain a 3rd work graph, and the 3rd work graph is input into a 7×7 Conv module to obtain a 4th work graph; The 1st segmented graph, the 1st work graph, the 2nd work graph, the 3rd work graph and the 4th work graph are spliced to obtain a 5th work graph, the 5th work graph is input into a 1×1 Conv module to obtain a 6th work graph, and the 6th work graph is taken as the output of the target C2F module.

5. The method of claim 1, wherein, Segmenting the target label graph according to the label to obtain a feature extraction graph includes: An area standard image is obtained, a work feature graph is obtained by processing the target label graph through a U-net network, a segmentation mask is obtained by convolving the work feature graph, a target mask is obtained by performing a preset operation on the segmentation mask, a target standard image is obtained by performing a preset operation on the area standard image, and a light feature graph is obtained by fusing the target mask and the target standard image; the preset operation includes a convolution operation and an up-sampling operation; The light feature graph and the work feature graph are fused to obtain a light enhancement graph, and the light enhancement graph is segmented according to the label to obtain a feature extraction graph.

6. An intelligent safety assessment system for a work trade operation scene, characterized in that, The system includes: An image extraction module is configured to acquire a video stream of a target region in real time, and frame the video stream to obtain a video image set; the video image set includes a plurality of video images; An image recognition module is configured to enhance a target video image to obtain a target enhanced image, recognize the target enhanced image through a target model to obtain a label, and add the label to the target enhanced image to obtain a target labeled image; the target video image is any one of the video images in the video image set; An occlusion judgment module is configured to segment the target labeled image according to the label to obtain a feature extraction image, and judge the feature extraction image to obtain a judgment result; the judgment result includes occlusion and no occlusion; A first judgment module is configured to, if the judgment result is no occlusion, perform safety evaluation on a target person to obtain an evaluation result, and if the evaluation result is risk, perform risk warning; A second judgment module is configured to, if the judgment result is occlusion, perform sound and light warning, and if it is still determined as occlusion within a preset time period, perform risk warning.

7. The system as claimed in claim 6, wherein, The target model is improved based on YOLOv8 and includes: The Conv modules in the 0th layer, the 1st layer, the 3rd layer, the 5th layer and the 7th layer in the YOLOv8 model are replaced with CBS modules, the neck structure of YOLOv8 is replaced with a target neck structure to obtain a target model; The working process of the target neck structure includes: The output of the 9th layer SPPF module is taken as the 1st input image of the target neck structure, the output of the 8th layer C2f module is taken as the 2nd input image of the target neck structure, the output of the 6th layer C2f module is taken as the 3rd input image of the target neck structure, the output of the 4th layer C2f module is taken as the 4th input image of the target neck structure, and the output of the 5th layer C2f module is taken as the 5th input image of the target neck structure; The 1st input image is input into the CBS module to obtain a 1st feature image, the 4th input image is input into the CBS module to obtain a 2nd feature image, the 1st feature image, the 2nd feature image and the 3rd input image are input into an adaptive fusion module to obtain a 3rd feature image, the 3rd feature image is sequentially input into a target C2F module and a CBS module to obtain a 4th feature image, the 5th input image is input into the CBS module to obtain a 5th feature image, and the 4th input image, the 4th feature image and the 5th feature image are input into the adaptive fusion module to obtain a 6th feature image; The 6th feature image is input into the C2f module to obtain a 7th feature image, the 2nd input image, the 3rd input image and the 4th input image are input into the adaptive fusion module to obtain an 8th feature image, and the 7th feature image and the 8th feature image are added to obtain a 9th feature image, which is taken as the input of a detection head; After upsampling the 9th feature image, the 9th feature image is spliced with the 5th input image to obtain a 10th feature image, the 10th feature image is added with the 1st input image to obtain an 11th feature image, and the 11th feature image is taken as the input of the detection head; The 6th feature map is input into a CBS module to obtain a 12th feature map, the 4th feature map and the 12th feature map are spliced to obtain a 13th feature map, the 13th feature map is sequentially input into a target C2F module and a CBS module to obtain a 14th feature map, the 1st feature map and the 14th feature map are spliced to obtain a 15th feature map, and the 15th feature map is taken as the input of the detection head. 8.The intelligent safety assessment system for a work site according to claim 7, wherein, The working principle of the adaptive fusion module comprises: An original feature tensor is obtained, and the original feature tensor is down-sampled to obtain a high-scale feature tensor, a medium-scale feature tensor and a low-scale feature tensor; The high-scale feature tensor and the medium-scale feature tensor are input into a multi-scale fusion module to obtain a 1st tensor, and the medium-scale feature tensor and the low-scale feature tensor are input into a multi-scale fusion module to obtain a 2nd tensor; The 1st tensor is input into a channel attention module to obtain a 3rd tensor, the 3rd tensor and the 2nd tensor are spliced to obtain a 4th tensor, the 4th tensor and the low-scale feature tensor are spliced to obtain a 5th tensor, and the 5th tensor is taken as the output of the adaptive fusion module. 9.The intelligent safety assessment system for a work site according to claim 7, wherein, The improvement of the target C2F module comprises: The bottleneck module in the C2F module is replaced by a target bottleneck module to obtain a target C2F module; The improvement of the target bottleneck module comprises: The Conv module in the bottleneck module is replaced by a target Conv module to obtain a target bottleneck module: The working principle of the target Conv module comprises: An initial work graph is obtained, the initial work graph is segmented to obtain a 1st segmented graph and a 2nd segmented graph, the 2nd segmented graph is input into a 1x1 Conv module to obtain a 1st work graph, the 1st work graph is input into a 3x3 Conv module to obtain a 2nd work graph, the 2nd work graph is input into a 5x5 Conv module to obtain a 3rd work graph, and the 3rd work graph is input into a 7x7 Conv module to obtain a 4th work graph; The 1st segmented graph, the 1st work graph, the 2nd work graph, the 3rd work graph and the 4th work graph are spliced to obtain a 5th work graph, the 5th work graph is input into a 1x1 Conv module to obtain a 6th work graph, and the 6th work graph is taken as the output of the target C2F module.

10. The system as claimed in claim 6, wherein, The occlusion judgment module comprises: A feature extraction module is configured to obtain a region standard image, process the target label graph through a U-net network to obtain a work feature graph, perform convolution on the work feature graph to obtain a segmentation mask, perform a preset operation on the segmentation mask to obtain a target mask, perform a preset operation on the region standard image to obtain a target standard image, and fuse the target mask and the target standard image to obtain an illumination feature graph; the preset operation comprises convolution operation and up-sampling operation. An image segmentation module is configured to fuse the illumination feature graph and the work feature graph to obtain an illumination enhanced graph, and segment the illumination enhanced graph according to a label to obtain a feature extraction graph.

Images