A method and system for identifying abnormal human behavior in tunnel construction environments

Abstract

This invention relates to the field of behavior detection technology and discloses a method and system for identifying abnormal human behavior in a tunnel construction environment. The method includes: inputting a tunnel construction image; performing low-light enhancement on the image using wavelet transform and causal inference based on a coordinate attention mechanism and multi-scale fusion mechanism to obtain a low-light enhanced image; processing and extracting features from the low-light enhanced image using an encoder with a multi-branch structure based on the content complexity of the image; generating dynamic attention weights based on a sharpness-aware dynamic attention mechanism for different areas of dust concentration in the low-light enhanced image, and processing accordingly based on the sharpness level; obtaining a dust-reduced and denoised image after decoding; inputting the dust-reduced and denoised image into a large visual model fine-tuned with domain knowledge of tunnel construction personnel images, and outputting the result of identifying abnormal human behavior in a tunnel construction environment. This invention can achieve accurate detection of abnormal behavior of construction personnel.

Description

Technical Field

[0001] This invention relates to the field of behavior detection technology, and in particular to a method and system for identifying abnormal human behavior in a tunnel construction environment. Background Technology

[0002] The tunnel construction environment is generally dark and complex, with mixed and frequently changing lighting, high levels of dust and noise. If a worker falls, it is difficult to detect, potentially causing significant losses. To ensure personnel safety, entry into blasting warning areas and around rotating machinery is prohibited. Furthermore, due to the numerous procedures and work groups involved in tunnel construction, overlapping operations could lead to fights between workers. Safety management regulations require workers to wear safety helmets and reflective vests, and prohibit smoking and dangerous behaviors such as dangling their feet near edges. Relying on manual inspections to detect these behaviors is costly and impractical for 24 / 7 monitoring. There is an urgent need for intelligent, round-the-clock monitoring methods to identify these abnormal behaviors, thereby improving tunnel safety management.

[0003] In recent years, object detection technology has made significant progress in the field of abnormal behavior monitoring. Researchers have effectively improved the detection accuracy and robustness in conventional scenarios by introducing attention mechanisms (such as SE and CBAM), optimizing feature fusion modules (such as PANet), and designing dedicated loss functions (such as WIoU) based on mainstream frameworks such as Faster R-CNN, DETR, DCGAN, YOLOv8, and YOLOv11. These methods demonstrate good performance in well-lit and clean controlled environments, but when transferred to the complex, dark, and dusty environment of tunnel construction, they face multiple deep-seated challenges due to the dual contradictions between scene characteristics and technology adaptability.

[0004] Artificial light sources in tunnel construction environments suffer from low illuminance (typically below 50 lux) and uneven distribution, resulting in images that are generally characterized by insufficient brightness, low contrast, and illumination distortion. This leads to blurred target outlines and weakened texture information, making it difficult for traditional feature extraction modules to separate effective features from noise interference. Simultaneously, the large amount of dust generated during construction not only reduces environmental visibility (less than 5 meters in some scenes) but also easily adheres to lenses, creating a fogging effect that blurs target edges and diffuses features, further exacerbating the difficulty of feature extraction. Existing mainstream algorithms lack targeted illumination correction and dehazing enhancement mechanisms; their shallow convolutional layers are easily contaminated by noise, and the representational ability of deep semantic features is significantly reduced due to input information distortion. More importantly, the dynamic suspension characteristics of dust cause randomly distributed bright / dark spots in the image, mixing with target features such as construction workers and equipment. Traditional threshold segmentation and feature selection mechanisms cannot effectively distinguish between them, directly causing target localization errors and misclassification.

[0005] Tunnel construction environments are dark, dusty, and prone to interference, while the feature extraction mechanisms of mainstream frameworks lack anti-interference design and are difficult to adapt to such scenarios. Faster R-CNN employs a fixed anchor design, where the scale and proportion of the anchor boxes depend on predefined assumptions. In scenarios where the tunnel target outline is blurred and the scale varies, the accuracy of anchor box matching drops significantly. Furthermore, its CNN-based feature extraction network has a weak ability to capture details in low signal-to-noise ratio images, and small targets are sparsely represented in deep feature maps, further exacerbating the risk of missed detections.

[0006] DETR relies on a global Transformer attention mechanism, and its computational complexity increases quadratically with the feature map size. It cannot handle high-resolution tunnel images to preserve small target features, and its position encoding is sensitive to illumination distortion. Feature alignment accuracy decreases significantly under low illumination, and small target detection is an inherent weakness of its design.

[0007] As a generative adversarial network, DCGAN's generator tends to learn texture features of general scenes, making it difficult to generate realistic degradation features in dusty tunnel environments. Furthermore, its training is prone to mode collapse, and it cannot provide diverse scene-specific augmentation samples, making it difficult for data augmentation strategies based on it to improve the feature quality of tunnel images.

[0008] While YOLOv8 and YOLOv11 are real-time detection frameworks, their downsampling operations in the backbone are prone to losing fine-grained features in low-light images. Small targets occupy only a tiny pixel area in the feature map, and the anchor-free design results in insufficient bounding box regression accuracy in scenes with blurred target edges. Furthermore, neither framework has built-in illumination correction and dehazing modules, shallow convolutional layers are easily contaminated by dust and noise, and the ability to represent deep semantic features is significantly reduced due to input information distortion. In addition, existing frameworks have not designed scene-specific loss functions for tunnel construction environments.

[0009] Furthermore, the unique characteristics of tunnel construction scenarios lead to a triple dilemma in dataset construction: quantity, quality, and adaptability. First, the hazardous nature of the construction environment and the continuity of operations limit the feasibility of data collection. Existing public datasets rarely cover typical features of tunnel scenarios such as low light, dust blurring, and dense crowd occlusion, resulting in a lack of sufficient scenario-based samples for model training. Second, the annotation process relies on professional knowledge in the construction domain, and the variable poses of targets in the tunnel environment (such as bending over to work) and frequent occlusion (such as material accumulation occlusion) not only make the annotation time for a single image 3-5 times longer than in conventional scenarios, but also easily lead to annotation ambiguity and errors, making the acquisition of high-quality annotated data extremely costly. Finally, there are significant domain differences between existing general datasets and tunnel scenarios. The feature distribution trained on general datasets (such as COCO and VOC) shows a significant domain shift compared to tunnel dust and low light scenarios, leading to a sharp decline in model generalization performance during cross-domain transfer.

[0010] Furthermore, due to the lack of annotation information for degradation features such as fogging and lighting distortion, the model struggles to learn scene-specific features during training, making it prone to overfitting and lacking robustness in practical applications. Summary of the Invention

[0011] To address the aforementioned problems, this invention proposes a method and system for identifying abnormal personnel behavior in tunnel construction environments, applicable to dimly lit and dusty tunnel construction scenarios. This invention employs low-light enhancement and dust / noise reduction technologies, combined with specialized knowledge of personnel behavior in tunnel construction environments to fine-tune a large-scale visual model, enabling accurate detection of abnormal personnel behavior and providing technical support for tunnel construction safety monitoring.

[0012] The technical solution adopted in this invention is as follows: A method for identifying abnormal human behavior in a tunnel construction environment, comprising: Dark light enhancement: Input a tunnel construction image, and perform dark light enhancement on the tunnel construction image through wavelet transform and causal inference based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a dark light enhanced image; Dust removal and noise reduction: Based on the content complexity of the low-light enhancement image, it is processed and features are extracted by an encoder with a multi-branch structure; then, based on the dust concentration in different areas of the low-light enhancement image, dynamic attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. After processing by the decoder, the dust removal and noise reduction image is obtained. Behavior recognition: The dust-reduced and noise-reduced image is input into a large visual model that has been fine-tuned with domain knowledge of tunnel construction personnel images. The model is then judged by a detection head, and the result of abnormal personnel behavior recognition in the tunnel construction environment is output.

[0013] Furthermore, the low-light enhancement process includes the following steps: The input tunnel construction image is sequentially processed by convolution and coordinate attention mechanism to extract a weighted feature map that contains both spatial detail location information and effective channel features; the weighted feature map is then subjected to wavelet transform to decompose it into one low-frequency component and three high-frequency components. The high-frequency and low-frequency components are enhanced separately, and the enhanced high-frequency and low-frequency components are fused using multi-scale features to obtain fused features. The semantic consistency of the fused features is controlled by causal reasoning, and then the fused features are reconstructed to obtain a low-light enhanced image.

[0014] Furthermore, the enhancement processing of the high-frequency components includes: concatenating the three high-frequency components along the channel dimension, adjusting the number of channels to match the input requirements; enhancing edge and texture details and suppressing noise pseudo-details; outputting the enhanced high-frequency feature map through residual connections and activation functions; splitting it into three high-frequency components by channel; and further processing each high-frequency component through lightweight convolutional blocks and attention gating mechanisms to obtain the enhanced high-frequency components.

[0015] Furthermore, the enhancement processing of the low-frequency components includes: employing parallel processing in both the spatial and frequency domains, wherein the spatial branch captures local illumination gradient patterns through convolution to preserve spatial continuity, and the frequency branch is transformed to the frequency domain via Fourier transform. After weighted enhancement of the low-frequency components related to the global brightness distribution using a learnable frequency domain mask, it is transformed back to the spatial domain via inverse Fourier transform. The features output from the spatial and frequency domain branches are fused by channel attention weighting, and then the number of channels is adjusted by convolution to obtain the enhanced low-frequency components.

[0016] Furthermore, the process of controlling the semantic consistency of the fused features through causal reasoning includes: constructing a causal graph from the low-light image to the enhanced image, identifying and blocking interference paths from noise to false details and from uneven lighting to semantic ambiguity; introducing semantic loss, extracting the visual features of the enhanced image and the semantic features of the text description of the normal lighting scene, and minimizing the cosine distance between the two.

[0017] Furthermore, the dust removal and noise reduction process includes the following steps: The richness of detail in image regions is quantified by gradient magnitude variance and edge density indices. The input image is divided into blocks of different sizes. Three parallel branches are used to process the blocks of different sizes respectively. The first branch extracts detail features, the second branch captures regional structural features, and the third branch captures global features. The output features of the three branches are fused by cross-scale attention weighting to obtain fused features; a weight map is generated based on the sharpness-aware dynamic attention mechanism to adjust the weight of the fused features; the weighted fused features are normalized and dimension-unified, and then processed by the decoder to obtain a dust-reduced and denoised image.

[0018] Furthermore, the generation of the weight map based on the clarity-aware dynamic attention mechanism includes: Features are extracted using a lightweight fully convolutional network to generate a blur map, where each pixel value in the blur map represents the relative blur level of the corresponding region; After standardizing the blur value to the range of 0-1 using the activation function, the weights are reversed to obtain the sharpness weight map. The clarity weight map is expanded to the same number of channels as the fusion feature, and then multiplied element-wise with the fusion feature to complete the feature weighting.

[0019] Furthermore, the behavior recognition includes the following steps: A large visual model pre-trained with general image samples is selected as the base model. Most of the pre-trained parameters of the base model are frozen, and adapters are inserted only in the last few layers. Low-rank matrices are inserted in the unfrozen projection layers. We used a dataset of tunnel construction worker behavior images containing bounding boxes and behavior labels, and combined an optimizer and a learning rate scheduling strategy of linear temperature rise plus cosine decay to fine-tune the large visual model. Based on the fine-tuned visual model, abnormal behavior is identified using a dual-branch detection head that combines target regression and behavior classification, and the results of abnormal personnel behavior identification in the tunnel construction environment are output.

[0020] Furthermore, the abnormal behavior identification using the dual-branch detection head of target regression and behavior classification includes: the target regression branch calculates the loss between the predicted bounding box and the labeled bounding box using the overlap loss function; the behavior classification branch performs global average pooling on the features of the personnel region, processes them through a fully connected network, and outputs a probability distribution containing normal behavior and various abnormal behaviors through an activation function, and sets confidence and probability thresholds respectively to achieve abnormal behavior judgment.

[0021] A system for identifying abnormal human behavior in a tunnel construction environment includes: The low-light enhancement module is configured to enhance the low-light image of the tunnel construction image by using wavelet transform and causal inference methods based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a low-light enhanced image. The dust removal and noise reduction module is configured to process and extract features from the content of the low-light enhancement image through an encoder with a multi-branch structure, based on the content complexity of the low-light enhancement image; then, based on the dust concentration in different areas of the low-light enhancement image, dynamic attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. After processing by the decoder, the dust removal and noise reduction image is obtained. The behavior recognition module is configured to input the dust-reduced and noise-reduced image into a large visual model finely adjusted by domain knowledge of tunnel construction personnel images, make judgments through the detection head, and output the recognition results of abnormal personnel behavior in the tunnel construction environment.

[0022] The beneficial effects of this invention are as follows: 1. This invention employs low-light enhancement and dust / noise reduction technologies, combined with specialized knowledge of personnel behavior in tunnel construction environments to fine-tune a large-scale visual model. This enables accurate identification of abnormal behaviors of construction workers in dimly lit and dusty tunnel construction environments. Compared to the baseline model YOLO11n, this invention improves the average precision by 7%, accuracy by 8%, and recall by 6% in abnormal behavior detection across all tunnel construction scenarios, thus enhancing the overall intelligence level of tunnel safety management.

[0023] 2. To address the challenges of low-light environments in tunnel construction and the lack of low-light adaptability in existing algorithms, this invention proposes a dark-light enhancement technique. By employing coordinate attention mechanisms, wavelet transform, cross-scale fusion, and causal inference, the brightness of low-light images is increased, resulting in clearer and more distortion-free images. In detailed low-light tunnel scenarios, by introducing this dark-light enhancement technique alone, compared to the baseline model YOLO11n, this invention achieves a 6% improvement in average precision, a 7% improvement in accuracy, and a 7% improvement in recall for anomaly detection across all tunnel construction scenarios, effectively alleviating the difficulty of low-light image recognition.

[0024] 3. To address the issue of dusty environments in tunnel construction and the lack of dust removal design in existing algorithms, this invention proposes a dust removal and noise reduction technique. This technique combines an image content complexity judgment mechanism, a multi-branch encoder, multi-scale fusion, and clarity-aware dynamic attention to effectively remove dust and reduce noise from images, improving image clarity. After introducing this dust removal and noise reduction technique, compared to the baseline model YOLO11n, this invention improves the average precision by 5%, accuracy by 6%, and recall by 5% in anomaly detection across all tunnel construction scenarios, thus contributing to improved detection accuracy in dusty tunnel construction environments.

[0025] 4. To address the issues of limited, difficult-to-obtain, and high-cost annotation datasets for tunnel construction, this invention utilizes a large visual model, fine-tuned with tunnel construction images, to enrich sample features. In a small-sample scenario using only 2000 labeled images, compared to the baseline model YOLO11n (with similar labeled data), this invention achieves a 3% improvement in average precision, a 5% improvement in accuracy, and a 4% improvement in recall. This effectively alleviates the problems of limited samples and high annotation costs in tunnel construction environments, achieving a breakthrough improvement in detection performance under small-sample conditions. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for identifying abnormal human behavior in a tunnel construction environment according to Embodiment 1 of the present invention.

[0027] Figure 2 This is a flowchart of a method for identifying abnormal human behavior in a tunnel construction environment according to Embodiment 3 of the present invention.

[0028] Figure 3 This is a flowchart of the dark light enhancement process in Embodiment 3 of the present invention.

[0029] Figure 4 This is a flowchart of dust removal and noise reduction in Embodiment 3 of the present invention. Detailed Implementation

[0030] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] Example 1 like Figure 1 As shown, this embodiment provides a method for identifying abnormal personnel behavior in a tunnel construction environment, including: Dark light enhancement: Input a tunnel construction image, and perform dark light enhancement on the tunnel construction image through wavelet transform and causal inference based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a dark light enhanced image; Dust removal and noise reduction: Based on the content complexity of the low-light enhancement image, a multi-branch encoder is used for processing and feature extraction; then, based on the dust concentration in different areas of the low-light enhancement image, attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. The image is then processed by the decoder to obtain the dust-reduced and noise-reduced image. Behavior recognition: Input the dust-reduced and noise-reduced image into the large visual model that has been fine-tuned with domain knowledge of tunnel construction personnel images, and make judgments through the detection head to output the results of abnormal personnel behavior recognition in the tunnel construction environment.

[0032] The method for identifying abnormal human behavior in this embodiment is described in detail below.

[0033] I. Dark Light Enhancement Due to the dim lighting environment of tunnel construction and the lack of low-light adaptation mechanisms in existing algorithms, coupled with image quality degradation, detection accuracy is insufficient. Specifically, the overall environment of tunnel construction is dark and complex, and low lighting conditions cause a decrease in detection accuracy. The low-light characteristics of the tunnel construction environment directly lead to image quality degradation, significantly reducing the contrast between the target and the background, and severely weakening key recognition features such as edge textures. At the same time, the low-light environment amplifies noise, causing noise to mix with target features and interfering with the model's feature extraction process. Existing mainstream frameworks such as Faster R-CNN and YOLOv8 / 11 do not have built-in lighting correction mechanisms, making it difficult to handle problems such as insufficient image brightness, low contrast, and noise mixing under low-light conditions (≤50 lux) in tunnels. This leads to feature extraction failure, positioning misalignment, and misjudgment, ultimately resulting in a significant decline in detection accuracy and failing to meet the precise requirements for monitoring abnormal behavior of tunnel construction personnel.

[0034] Based on this, this embodiment proposes a low-light enhancement strategy based on wavelet transform, the main process of which is as follows: convolutional layer extracts image features; coordinate attention mechanism is used for weight allocation; wavelet transform is performed on the input low-light image to decompose it into low-frequency components and three high-frequency components (horizontal, vertical and diagonal); feature extraction further mines and optimizes the features; high-frequency components are processed by high-frequency enhancement, and high-frequency Mamba modules are used to enhance details; low-frequency components are processed by low-frequency enhancement, and fast Fourier convolution (FFC) is used to enhance global features; multi-scale feature fusion; CLIP semantic loss is used to ensure the semantic consistency of the enhanced image; and upsampling is used to reconstruct the enhanced image.

[0035] II. Dust Removal and Noise Reduction Due to the dusty environment of tunnel construction and the lack of dedicated dust and fog removal design in existing algorithms, feature extraction becomes difficult due to dust and fogging blurring. Specifically, core processes in tunnel construction such as blasting, slag removal, and shotcreting continuously generate large amounts of dust, keeping the environmental dust concentration at a high level for extended periods. The pervasive dust particles significantly reduce visibility. These suspended dust particles strongly scatter light, weakening the efficiency of light signal transmission. Simultaneously, fine particles easily accumulate on the lens surface of detection equipment, forming an optical scattering layer and causing fogging blurring. This dual interference leads to a significant decrease in the contrast of the acquired images, severely obscuring target edges and texture details, resulting in a substantial reduction in the effective visual information received. Existing algorithms such as Faster R-CNN and YOLOv8 / 11 lack specific dust removal mechanisms for tunnel scenarios and cannot address issues such as reduced image contrast caused by dust scattering and lens fogging, target detail obscuration, and feature diffusion. This makes it difficult to distinguish targets from the background, leading to increased localization errors and higher classification misclassification rates, ultimately resulting in a comprehensive decline in detection performance and failing to meet the precise requirements for monitoring abnormal behavior of tunnel construction personnel.

[0036] Based on this, this embodiment proposes a multi-branch attention-driven dust removal and noise reduction strategy. The main process is as follows: First, the input image is divided into multi-scale blocks according to the complexity of the image content; a multi-branch encoder structure is designed, including a multi-layered Swintransformer structure; a multi-scale feature fusion module is designed to fuse features; a sharpness-aware dynamic attention module is designed to generate corresponding weighted attention based on the sharpness of different regions; a feature adaptation module is designed to normalize features; a decoder structure containing multi-layered Swintransformer blocks is designed, employing a cross-stage feature fusion mechanism; finally, the dust-removed image is output.

[0037] III. Behavior Recognition Existing algorithms rely on large-scale labeled data, but tunnel scene datasets are scarce and labeling costs are high. Specifically, existing algorithms have an enormous demand for labeled data, while the unique environment of tunnel construction presents significant challenges to data labeling. Publicly available datasets are extremely limited, and data acquisition is difficult. Dim lighting and dusty environments blur image details, and frequent occlusion by construction personnel, equipment, and materials further increases the difficulty of labeling. Acquiring large amounts of data and completing labeling requires significant manpower, and the massive workload may also make it difficult to guarantee labeling accuracy.

[0038] Based on this, this embodiment proposes an optimization scheme based on the visual large model DINOv3. By using a small amount of labeled data on personnel behavior in the tunnel construction environment for LoRA fine-tuning, the problems of difficult data acquisition and huge labeling workload are solved, and the accurate detection of abnormal personnel behavior in the tunnel construction environment is achieved.

[0039] In summary, this method can accurately identify and trigger alarms for abnormal behaviors in the dim and dusty environment of tunnel construction (such as personnel falling, fighting, not wearing protective equipment as required, entering restricted areas, smoking, and feet dangling in the air near the edge), which helps to improve the level of intelligence in tunnel safety management.

[0040] Accordingly, this embodiment also provides a system for identifying abnormal personnel behavior in a tunnel construction environment, including: The low-light enhancement module is configured to enhance the low-light image of the tunnel construction image by using wavelet transform and causal inference methods based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a low-light enhanced image. The dust removal and noise reduction module is configured to process and extract features from the content of the low-light enhancement image through an encoder with a multi-branch structure, based on the content complexity of the low-light enhancement image; then, based on the dust concentration in different areas of the low-light enhancement image, dynamic attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. After processing by the decoder, the dust removal and noise reduction image is obtained. The behavior recognition module is configured to input the dust-reduced and noise-reduced image into a large visual model finely adjusted by domain knowledge of tunnel construction personnel images, make judgments through the detection head, and output the recognition results of abnormal personnel behavior in the tunnel construction environment.

[0041] Example 2 This embodiment provides a method for identifying abnormal human behavior in a tunnel construction environment, including: given an image input, firstly, using a low-light enhancement module to enhance the low-light performance of the tunnel construction image; then, using a dust removal and noise reduction module to remove dust and reduce noise in the image to improve image clarity; after completing the low-light enhancement and dust removal and noise reduction, the image input is tested using a large visual model finely tuned with professional domain knowledge and LORA; finally, the result of identifying abnormal human behavior in a tunnel construction environment is output.

[0042] Specifically, firstly, after image input, a low-light enhancement module is used to enhance the low-light performance of the tunnel construction image. This is achieved through wavelet transform and causal inference, incorporating coordinate attention and multi-scale fusion mechanisms, thereby improving detection accuracy. Secondly, a dust removal and noise reduction module is used to remove dust and noise from the image. Based on the complexity of the image content, an encoder with a multi-branch structure is used to process the image and extract features. Then, for different dust concentrations in different areas of the image, a sharpness-aware dynamic attention module is employed to generate dynamic attention weights, which are then processed accordingly based on the sharpness level. After processing by the decoder, the dust-removed and noise-reduced image is output. Finally, a large-scale visual model, refined using domain knowledge of tunnel construction personnel images and LORA fine-tuning, is used by the detection head to detect abnormal behavior of personnel in the tunnel construction environment.

[0043] The method for identifying abnormal human behavior in this embodiment is described in detail below.

[0044] I. Low Light Enhancement Module Due to limited lighting and a generally dim and complex environment during tunnel construction, target information cannot be clearly identified. To address this issue, this embodiment designs a low-light enhancement module to improve detection performance. The main structure of this module is as follows.

[0045] 1. The input image undergoes feature extraction via convolution and coordinate attention mechanisms. Preferably, the image is first processed through convolutional layers to output a feature map; subsequently, the feature map enters the coordinate attention mechanism unit, where global pooling is performed in both the horizontal and vertical directions to capture spatial location dependencies, and then channel attention weights are learned through two independent fully connected layers; the final output weighted feature map retains spatial detail and location information while highlighting effective features in the channel dimensions, laying the foundation for subsequent processing. Preferably, the calculation method of the coordinate attention mechanism is as follows:

[0046] in, It is the output feature. These are input features. It is the vertical attention weight. It is the horizontal attention weight.

[0047] 2. Wavelet transform for frequency domain feature separation. Preferably, wavelet transform decomposes the feature map after coordinate attention processing into one low-frequency component and three high-frequency components. The low-frequency component reflects the basic structure of the image, such as the global illumination distribution and smooth regions, and is the main component responsible for the insufficient brightness of low-light images. The high-frequency components include horizontal, vertical, and diagonal directions, corresponding to edge and texture details in each direction, respectively, and are responsible for the blurred details and significant noise in low-light images. The decomposition process is completed through convolution operations, ultimately yielding four frequency sub-bands.

[0048]

[0049] Where L represents the low-frequency component obtained after wavelet transform decomposition, H represents the transformed horizontal high-frequency component, V represents the transformed vertical high-frequency component, and D represents the transformed diagonal high-frequency component; WT represents the wavelet transform, I dark This represents the input low-light image.

[0050]

[0051] Among them, L fea H represents the output characteristics of the low-frequency components after wavelet convolution processing. hH represents the output features of the horizontal high-frequency components after feature extraction. v H represents the output features of the vertical high-frequency components after feature extraction. d This represents the output features of the diagonal high-frequency components after feature extraction. DW represents wavelet convolution. This represents depthwise separable convolution. Represents horizontal convolution. Represents vertical convolution. This represents diagonal convolution.

[0052] 3. High-Frequency Enhancement. Preferably, the three high-frequency feature maps—horizontal, vertical, and diagonal—are first concatenated along the channel dimension, and then the number of channels is adjusted to match the input requirements of the high-frequency Mamba unit. The high-frequency Mamba unit achieves efficient modeling of detail sequences based on a state-space model. It first extracts local detail patterns through depthwise separable convolution, then uses a selective state update mechanism to enhance key details such as edges and textures while suppressing noise pseudo-details. Finally, it outputs the enhanced high-frequency feature map through residual connections and the GELU activation function, and splits it into three processed high-frequency components by channel. Each component is then further processed by a lightweight convolutional block to enhance edge continuity and suppress residual noise through an attention gating mechanism, ensuring accurate and appropriate detail enhancement. Specifically, the relevant formulas are as follows:

[0053]

[0054]

[0055] in, Indicates horizontal selective scanning. Representation of features Output results after horizontally selective scanning enhancement; Indicates vertical selective scanning. Representation of features Output results after vertical selective scanning enhancement; Indicates diagonal selective scanning. Representation of features Output results after diagonal selective scanning enhancement.

[0056] 4. Low-frequency enhancement. Preferably, the low-frequency feature map is divided into spatial and frequency domain branches for parallel processing using Fast Fourier Transform (FFT). The spatial branch captures local illumination gradient patterns through convolution to preserve spatial continuity, while the frequency branch transforms the feature map to the frequency domain using FFT. A learnable frequency domain mask is then used to weight and enhance low-frequency components such as global brightness distribution. Finally, an Inverse Fourier Transform is used to transform the feature map back to the spatial domain. The output features from both branches are then fused using channel attention weighting, with a focus on enhancing the brightness of low-light areas. The fused feature map is then adjusted for the number of channels using a 1×1 convolution, ultimately outputting low-frequency components with uniform illumination and clear structure.

[0057]

[0058] in, This represents the reconstructed image after inverse wavelet transform. This represents the inverse wavelet transform.

[0059] 5. Multi-scale feature fusion for cross-resolution information complementarity. Preferably, to further enhance the complementarity of multi-scale features, a cross-scale feature fusion mechanism is introduced to achieve deep integration of features at different resolutions. This step receives feature inputs from two key scales: first, high-resolution detail features optimized by a coordinate attention mechanism; and second, features with both local details and preliminary enhancements after the first-stage hierarchical feature restoration processing. Through operations such as feature alignment, channel dimension splicing, and attention weighting, the information limitations of a single scale can be broken, and the edge texture details of high-resolution features and the global illumination structure of low-resolution features can be semantically and spatially precisely fused, providing richer multi-scale contextual information for subsequent feature extraction, high-frequency enhancement, and low-frequency enhancement.

[0060] 6. Causal inference ensures the semantic consistency of the enhanced image. Preferably, a causal graph from the low-light image to the enhanced image is constructed to identify and block interfering paths such as noise leading to false details and uneven lighting leading to semantic ambiguity. Causal intervention forces the model to rely only on the causal path of real edges and textures when enhancing details, suppressing false causal associations caused by noise. At the same time, CLIP semantic loss is introduced. Utilizing the image-text semantic alignment capability of the CLIP model, the visual features of the enhanced image and the semantic features of the text description of the normal lighting scene are extracted, and the cosine distance between the two is minimized to ensure that the enhanced image is semantically consistent with the human perception of a normal lighting scene.

[0061] II. Dust Removal and Noise Reduction Module During tunnel construction, processes such as blasting, muck removal, and shotcreting can generate dust concentrations exceeding 8 mg / m³. The tunnel environment is characterized by persistently low visibility, and dust particles easily adhere to lenses, creating a fog-like blur. To ensure safe monitoring during construction, efficient dust removal of images is necessary to improve subsequent detection performance. The main structure of the dust removal and noise reduction module is as follows.

[0062] 1. Content Complexity Awareness-Based Adaptive Multi-Scale Block Partitioning Unit. Preferably, dynamic size matching is achieved by quantifying complexity through texture, edges, and other metrics, considering the differences between complex texture regions (such as construction machinery and densely populated areas) and simple texture regions (such as tunnel walls and open areas) in tunnel images. Specifically, gradient magnitude variance and edge density indices are used to quantify the richness of regional details, dividing the input image into blocks of three sizes. Small blocks are used for high-complexity regions to preserve fine structures, medium blocks balance details and global information, and large blocks focus on global scene representation.

[0063] 2. Multi-branch encoder structure. Preferably, three parallel branches are designed, each based on a multi-layer feature extraction structure built on the SwinTransformerV2 model. Branch 1 processes small blocks, extracting detailed features such as edges and textures through PatchEmbedding and multiple layers of SwinTransformerBlock; Branch 2 processes medium blocks, capturing region-level structural features; Branch 3 processes large blocks, capturing global features with the help of a large receptive field. The principle of the Transformer's attention mechanism is as follows:

[0064] in, This represents the attention mechanism. Q Represents the query vector. K Represents the key vector. V Represents a value vector; Represents the normalized exponential function, This represents the interaction between the query vector and the key vector. This represents the dimension of the key vector K.

[0065] 3. Multi-scale feature fusion unit. Preferably, the output features of the three branches are fused using cross-scale attention weighting. First, adaptive pooling maps the features of different branches to the same spatial dimension. Then, a multi-scale attention mechanism dynamically assigns weights to the detailed features of the high-complexity branch, the structural features of the medium-complexity branch, and the global features of the low-complexity branch. Finally, 1×1 convolution and feature concatenation are used to achieve efficient aggregation of multi-scale features, solving the scale bias problem of single-branch features and providing multi-scale feature support for dust density perception.

[0066] 4. Clarity-aware dynamic attention unit. Preferably, this unit includes the following structure: Sharpness Feature Extraction Layer: This layer receives feature maps from multi-scale fusion as input. It employs a lightweight fully convolutional network structure. First, convolutional layers reduce the number of channels and extract local texture and edge features. Then, a ReLU activation function is used to introduce non-linearity. Next, a 1×1 convolutional layer compresses the features into a single-channel output, generating a blur map where each pixel value represents the relative blur level of the corresponding region.

[0067] Dynamic Weight Generation Layer: Taking the blur map as input, its core task is to transform blur information into effective attention weights. The blur values ​​are standardized to the range of 0-1 using the Sigmoid function, and then a weight inversion operation is performed, giving low weights to high-blur regions and high weights to low-blur regions. This ensures the network focuses more on information-rich, sharp regions, whose features are more reliable and can effectively guide the blur removal process. The output of this layer is a sharpness weight map used to dynamically adjust feature importance.

[0068] Feature Weighting and Output Layer: This layer applies sharpness weights to the original features, completing the final attention adjustment. The original multi-scale fused features and the sharpness weight map are combined in this layer. A broadcast mechanism expands the weights to the same number of channels as the feature map, and then element-wise multiplication is performed, scaling each feature value according to the sharpness of its region. The weighted features retain all spatial and channel information, emphasizing features from sharp regions and minimizing the contribution from blurry regions, thus optimizing the allocation of computational resources.

[0069] 5. Feature Adaptation Unit and Decoder Structure. Preferably, the feature adaptation unit uses an RMSnorm layer to normalize features, eliminating feature distribution shifts caused by dust concentration and texture complexity in different blocks. Then, a linear transformation layer maps the features to a unified dimension, providing standardized input to the decoder. The decoder constructs a multi-layer structure based on SwinTransformerBlock, gradually recovering the high-resolution image through upsampling and the multi-layer structure. Simultaneously, a cross-stage feature fusion mechanism is introduced, stitching together the shallow detail features of each branch of the encoder with the deep global features of the decoder, solving the detail loss problem caused by simple upsampling. The final output is a clear tunnel image after dust removal, providing high visibility and high detail visual input for downstream tasks such as abnormal human behavior detection and equipment status recognition.

[0070] III. Specialized domain fine-tuning based on large visual models Detecting abnormal human behavior in tunnel construction environments is constrained by both small sample sizes and high annotation costs. Traditional models generally rely on large-scale labeled samples for full parameter training, requiring at least thousands or even tens of thousands of labeled data points to build a stable feature learning system. However, currently, datasets for human behavior detection in tunnel environments are extremely scarce, difficult to obtain, and costly to manually annotate. Therefore, this embodiment designs a method based on a large visual model and fine-tuned using domain-specific knowledge.

[0071] 1. Selection Logic and Core Advantages of the Base Model. Preferably, to overcome the challenge of small-sample adaptation, this embodiment selects DINOv3 as the base model, leveraging its powerful pre-trained general feature extraction capabilities and flexible adaptation potential. DINOv3 has been pre-trained on 1 billion general image samples, covering a large amount of diverse data such as human behavior and industrial scenes, and has deeply mastered core capabilities such as human contour recognition, behavior pattern abstraction, and scene semantic understanding. Compared to traditional models, DINOv3 does not require retraining based on large-scale tunnel scene-specific data. With the general capabilities accumulated during the pre-training stage, it can capture the unique features of tunnel scenes under small-sample conditions. This embodiment adopts a general feature + scene fine-tuning approach to solve the dependence of traditional models on massive labeled data.

[0072] 2. LoRA-based small-sample fine-tuning. Preferably, by collecting a small number of images of personnel behavior in tunnel construction environments, LoRA technology is used to perform lightweight fine-tuning of DINOv3. The core idea is to freeze the backbone parameters and precisely fine-tune key layers. That is, most of the pre-trained parameters of DINOv3 are frozen, and LoRA adapters are inserted only in the last few layers. Low-rank matrices are inserted into the unfrozen Q / V projection layers, and the weights of the original linear layers are decomposed into two low-rank matrices. The adaptation effect and computational cost are balanced by controlling the rank and scaling factor. The dataset is divided into training and test sets to ensure that the test set covers all abnormal behavior types. At the same time, the annotation content is simplified, requiring only the annotation of personnel bounding boxes and behavior labels, reducing annotation time and cost. The optimizer AdamW is selected, with a learning rate scheduling strategy of linear temperature rise and cosine decay.

[0073] 3. Dual-branch detection head setup. Preferably, to achieve accurate output of personnel localization and abnormal behavior judgment, this embodiment designs a dual-branch detection head of target regression + behavior classification. The target regression branch focuses on personnel localization, using the CIoU loss function to calculate the overlap loss between the predicted bounding box and the labeled bounding box. The behavior classification branch is responsible for determining the behavior type. It performs global average pooling on the personnel region features output by the target detection branch to obtain a fixed-dimensional vector, which is then processed through two fully connected layers. Finally, it outputs a probability distribution containing normal behavior and various abnormal behaviors through Softmax activation. This branch uses the cross-entropy loss function. Confidence and probability thresholds are set on both branches to ensure the accuracy and reliability of abnormal behavior detection.

[0074] IV. Loss Function In this embodiment, the total loss function consists of four sub-loss functions: L1 norm loss for dark light enhancement, perceptual loss for dust removal and noise reduction, CIoU loss for target regression, and cross-entropy loss for classification.

[0075] Preferably, the L1 norm loss is calculated as follows:

[0076] Where N represents the total number of samples involved in the calculation. This represents the true value of the i-th sample. This represents the predicted value of the i-th sample.

[0077] Preferably, perceived loss The calculation method is as follows:

[0078] in, This represents the number of channels in the feature map of the j-th layer. This represents the spatial height of the feature map at layer j. This represents the spatial width of the feature map at layer j. This represents the j-th layer feature extracted from the real image. This represents the j-th layer feature extracted from the predicted image.

[0079] Preferably, CIoU loss The calculation method is as follows:

[0080] Where IoU represents the intersection-union ratio, and ρ is a function for calculating the Euclidean distance. This indicates the coordinates of the center point of the prediction box. α represents the coordinates of the center point of the true bounding box, c represents the diagonal length of the smallest bounding rectangle, α represents the weight coefficient, and v represents the aspect ratio difference measure.

[0081] Preferably, multi-class cross-entropy loss The calculation method is as follows:

[0082] Where N represents the total number of samples involved in the calculation, and C represents the total number of categories. Indicates the true label, This represents the probability predicted by the model.

[0083] Preferably, the total loss function The calculation method is as follows:

[0084] in, , , , All are weighting factors.

[0085] Example 3 This embodiment is based on embodiment 2: like Figure 2 As shown, this embodiment provides a method for identifying abnormal personnel behavior in a tunnel construction environment, including: First, after image input, a dark light enhancement module is used to enhance the dark light of the tunnel construction image. Specifically, wavelet transform and causal inference with coordinate attention and multi-scale fusion mechanisms are used to enhance the dark light of the image, thereby improving the detection effect. Second, a dust removal and noise reduction module is used to remove dust and noise from the image. Based on the complexity of the image content, an encoder with a multi-branch structure is used to process the image and extract features. For different dust concentrations in different areas of the image, a sharpness-aware dynamic attention module is used to generate dynamic attention weights, and corresponding processing is performed according to sharpness. After processing by the decoder, the dust-removed and noise-reduced image is output. Finally, a large visual model fine-tuned with domain knowledge of tunnel construction personnel images and LORA is used to detect abnormal personnel behavior in the tunnel construction environment by a detection head.

[0086] like Figure 3As shown, the low-light enhancement module mainly performs the following steps: convolution to extract image features, weight allocation based on coordinate attention mechanism; wavelet transform: decompose the input low-light image into low-frequency components and three high-frequency components (horizontal, vertical, and diagonal) using wavelet transform; feature extraction is further mined and combined for optimization; high-frequency enhancement: process high-frequency components through high-frequency enhancement blocks, and enhance details using high-frequency Mamba units; low-frequency enhancement: process low-frequency components through low-frequency enhancement blocks, and enhance global features using fast Fourier convolution (FFC); multi-level feature extraction and recovery; multi-scale feature fusion; causal inference: ensure the semantic consistency of the enhanced image by utilizing causal intervention and CLIP semantic loss; and reconstruct the enhanced image through upsampling.

[0087] like Figure 4 As shown, the dust removal and noise reduction module mainly performs the following steps: First, according to the complexity of the image content, the input image is divided into multi-scale blocks, namely 64×64, 128×128, and 256×256; a multi-branch encoder structure is designed, including a multi-layered Swintransformer structure; a multi-scale feature fusion unit is designed to fuse features; a sharpness-aware dynamic attention unit is designed to generate corresponding weighted attention based on the sharpness of different regions; a feature adaptation unit is designed to normalize features; a decoder structure containing multiple layers of Swintransformer blocks is designed, employing a cross-stage feature fusion mechanism; finally, the image after dust removal and noise reduction is output.

[0088] Specifically, this embodiment collected 2000 images of tunnel construction workers, labeled the personnel detection boxes and classified abnormal behaviors, and set the training set and test set at an 8:2 ratio. LabelMe was used for labeling, and the images were divided into 7 categories: normal, falling, fighting, not wearing protective equipment as required, entering a restricted area, smoking, and feet dangling in the air near the edge.

[0089] The mean average precision (mAP) of each category, with an IoU threshold of 0.5, is used as the comprehensive evaluation index for detection accuracy. Precision (P) and recall (R) are selected as the evaluation indexes for model refinement.

[0090]

[0091]

[0092] Where AP represents average precision, TP represents true positives, FP represents false positives, and FN represents false negatives.

[0093] All experiments were conducted on an NVIDIA RTX A6000 based on the PyTorch framework. YOLO11n was used as the baseline for fair comparison. Input images were consistently 640×640 pixels. All models were optimized using the AdamW optimizer with an initial learning rate of 0.01, a batch size of 32, and a total of 500 epochs. Default data augmentation was used to improve the models' generalization ability.

[0094] Implementation Details of LoRA-Based Lightweight Fine-Tuning for Small Samples. This embodiment constructs a dataset of 2000 images for personnel behavior recognition in tunnel construction scenarios and uses LoRA technology to lightweightly adapt the DINOv3 model. The core of this method is to keep the core parameters of the model frozen and only fine-tune key layers. Specifically, this embodiment introduces a LoRA adapter into the Q and V projection layers of the attention module in the last 6 layers of the TransformerBlock in DINOv3, while keeping the remaining pre-trained parameters fixed. LoRA adapts to specific scenarios by decomposing the original weight matrix into the product of two low-rank matrices, minimizing computational burden. In this experiment, the rank is set to 16 and the scaling factor to 32 to balance model expressiveness and training efficiency. The 2000 images are divided into training and test sets in an 8:2 ratio. The test set covers all abnormal behavior categories to be identified to ensure comprehensive evaluation. The annotation information only includes personnel detection boxes and corresponding behavior labels, simplifying the annotation process and reducing costs. The AdamW optimizer was used during the training phase, with a batch size of 32. The learning rate scheduling adopted a combination of linear warm-up and cosine decay strategies, and a total of 500 training rounds were performed.

[0095] Design and Collaborative Working Mechanism of the Dual-Branch Detection Head. To simultaneously complete personnel localization and behavior state recognition, this embodiment designs a dual-branch detection head that performs target regression and behavior classification in parallel. The target regression branch is responsible for predicting the bounding box of the personnel location, using CIoU loss as the localization loss function with a weight of 0.5 to improve the matching accuracy of the detection box. The behavior classification branch performs global average pooling on the personnel region features extracted by the regression branch, and then passes them through two fully connected layers and a Softmax activation function to output the probability distribution of each behavior category. This branch uses the cross-entropy loss function with a weight of 0.5 to ensure balanced optimization of localization and classification tasks during training. The collaborative inference mechanism of the two branches is as follows: First, personnel targets with a detection confidence of not less than 0.7 are selected, and the highest probability category output by their classification branch is taken as the behavior recognition result; if the category belongs to abnormal behavior and its probability value exceeds 0.8, an alarm process is triggered. This mechanism effectively controls the false alarm rate while ensuring detection reliability.

[0096] The total loss function consists of four sub-loss functions: L1 loss for low-light enhancement, perceptual loss for dust removal and noise reduction, CIoU loss for object regression, and cross-entropy loss for classification. The weighting factors λ1, λ2, and λ3 are also present. 2、 λ 3、 λ4 is 0.2, 0.2, 0.3, and 0.3 respectively.

[0097] Experimental results show that this method can accurately identify abnormal behaviors (such as people falling, fighting, not wearing protective equipment as required, entering restricted areas, smoking, and feet dangling in the air near the edge) in dimly lit and dusty tunnel construction environments. Compared with the baseline model YOLO11n, this method can improve the average accuracy of abnormal behavior detection by 7%, the precision by 8%, and the recall by 6% in all tunnel construction scenarios, thus improving the overall level of intelligent tunnel safety management.

[0098] Example 4 This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement a method for identifying abnormal personnel behavior in a tunnel construction environment, as described in Embodiment 1 or 2. The computer program can be in the form of source code, object code, executable file, or some intermediate form.

[0099] Example 5 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for identifying abnormal personnel behavior in a tunnel construction environment, as described in Embodiment 1 or 2. The computer program can be in the form of source code, object code, executable file, or some intermediate form. The storage medium includes any entity or device capable of carrying computer program code, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0100] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0101] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A method for identifying abnormal human behavior in a tunnel construction environment, characterized in that, include: Dark light enhancement: Input a tunnel construction image, and perform dark light enhancement on the tunnel construction image through wavelet transform and causal inference based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a dark light enhanced image; Dust removal and noise reduction: Based on the content complexity of the low-light enhancement image, it is processed and features are extracted by an encoder with a multi-branch structure; then, based on the dust concentration in different areas of the low-light enhancement image, dynamic attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. After processing by the decoder, the dust removal and noise reduction image is obtained. Behavior recognition: The dust-reduced and noise-reduced image is input into a large visual model that has been fine-tuned with domain knowledge of tunnel construction personnel images. The model is then judged by a detection head, and the result of abnormal personnel behavior recognition in the tunnel construction environment is output. The low-light enhancement process includes the following steps: The input tunnel construction image is sequentially processed by convolution and coordinate attention mechanism to extract a weighted feature map that contains both spatial detail location information and effective channel features; the weighted feature map is then subjected to wavelet transform to decompose it into one low-frequency component and three high-frequency components. The high-frequency and low-frequency components are enhanced separately, and the enhanced high-frequency and low-frequency components are fused using multi-scale features to obtain fused features. The semantic consistency of the fused features is controlled by causal reasoning, and then the fused features are reconstructed to obtain a low-light enhanced image; The enhancement processing of the high-frequency components includes: concatenating the three high-frequency components along the channel dimension, adjusting the number of channels to match the input requirements; enhancing edge and texture details and suppressing noise pseudo-details; outputting the enhanced high-frequency feature map through residual connections and activation functions; splitting the high-frequency components into three high-frequency components according to channels; and further processing each high-frequency component through lightweight convolutional blocks and attention gating mechanisms to obtain the enhanced high-frequency components. The enhancement processing of the low-frequency components includes: parallel processing in both the spatial and frequency domains. The spatial branch captures local illumination gradient patterns through convolution to preserve spatial continuity, while the frequency branch is transformed to the frequency domain via Fourier transform. After weighted enhancement of low-frequency components related to the global brightness distribution using a learnable frequency domain mask, it is transformed back to the spatial domain via inverse Fourier transform. The features output from the spatial and frequency domain branches are fused by channel attention weighting, and then the number of channels is adjusted by convolution to obtain the enhanced low-frequency components. The process of controlling the semantic consistency of the fused features through causal reasoning includes: constructing a causal graph from the low-light image to the enhanced image, identifying and blocking interference paths from noise to false details and from uneven lighting to semantic ambiguity; introducing semantic loss, extracting the visual features of the enhanced image and the semantic features of the text description of the normal lighting scene, and minimizing the cosine distance between the two. The dust removal and noise reduction process includes the following steps: The richness of detail in image regions is quantified by gradient magnitude variance and edge density indices. The input image is divided into blocks of different sizes. Three parallel branches are used to process the blocks of different sizes respectively. The first branch extracts detail features, the second branch captures regional structural features, and the third branch captures global features. The output features of the three branches are fused by cross-scale attention weighting to obtain fused features; a weight map is generated based on the sharpness-aware dynamic attention mechanism to adjust the weight of the fused features; the weighted fused features are normalized and dimension-unified, and then processed by the decoder to obtain a dust-reduced and denoised image; The weight map generation based on the clarity-aware dynamic attention mechanism includes: Features are extracted using a lightweight fully convolutional network to generate a blur map, where each pixel value in the blur map represents the relative blur level of the corresponding region; After standardizing the blur value to the range of 0-1 using the activation function, the weights are reversed to obtain the sharpness weight map. The clarity weight map is expanded to the same number of channels as the fusion feature, and then multiplied element-wise with the fusion feature to complete the feature weighting.

2. The method for identifying abnormal personnel behavior in a tunnel construction environment according to claim 1, characterized in that, The behavior recognition includes the following steps: A large visual model pre-trained with general image samples is selected as the base model. Most of the pre-trained parameters of the base model are frozen, and adapters are inserted only in the last few layers. Low-rank matrices are inserted in the unfrozen projection layers. We used a dataset of tunnel construction worker behavior images containing bounding boxes and behavior labels, and combined an optimizer and a learning rate scheduling strategy of linear temperature rise plus cosine decay to fine-tune the large visual model. Based on the fine-tuned visual model, abnormal behavior is identified using a dual-branch detection head that combines target regression and behavior classification, and the results of abnormal personnel behavior identification in the tunnel construction environment are output.

3. The method for identifying abnormal personnel behavior in a tunnel construction environment according to claim 2, characterized in that, The method of identifying abnormal behavior using a dual-branch detection head that combines target regression and behavior classification includes: the target regression branch calculates the loss between the predicted bounding box and the labeled bounding box using an overlap loss function; the behavior classification branch performs global average pooling on the features of the personnel region, processes them through a fully connected network, and outputs a probability distribution containing normal behavior and various abnormal behaviors through an activation function, setting confidence and probability thresholds respectively to achieve abnormal behavior judgment.

4. A system for identifying abnormal personnel behavior in a tunnel construction environment, employing the method for identifying abnormal personnel behavior in a tunnel construction environment as described in claim 1, characterized in that... include: The low-light enhancement module is configured to enhance the low-light image of the tunnel construction image by using wavelet transform and causal inference methods based on coordinate attention mechanism and multi-scale fusion mechanism to obtain a low-light enhanced image. The dust removal and noise reduction module is configured to process and extract features from the content of the low-light enhancement image through an encoder with a multi-branch structure, based on the content complexity of the low-light enhancement image; then, based on the dust concentration in different areas of the low-light enhancement image, dynamic attention weights are generated according to the sharpness perception dynamic attention mechanism, and corresponding processing is performed according to the sharpness. After processing by the decoder, the dust removal and noise reduction image is obtained. The behavior recognition module is configured to input the dust-reduced and noise-reduced image into a large visual model finely adjusted by domain knowledge of tunnel construction personnel images, make judgments through the detection head, and output the recognition results of abnormal personnel behavior in the tunnel construction environment.

Images