Tunnel dust environment construction robot lightweight target detection method and system

Abstract

The application belongs to the technical field of target detection, and discloses a tunnel dim-dust environment construction robot lightweight target detection method and system, which comprises the following steps: acquiring a tunnel construction environment image and preprocessing; inputting the preprocessed image into a lightweight target detection network for target recognition, wherein the lightweight target detection network comprises a backbone network, a neck network and a detection head; a dynamic asymmetric perception module is arranged in the backbone network, which is used for extracting deep features and shallow features of the image through an asymmetric convolution kernel; a cross-complementary fusion module is arranged in the neck network, which is used for cross-fusion of the deep features and the shallow features; a multiple pooling adaptive exposure module is arranged in the detection head, which is used for adaptive light compensation of the fused features, and target recognition based on the light-compensated features. The application can reduce the model complexity while improving the detection accuracy and robustness in dim and high-dust environments.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a lightweight target detection method and system for construction robots in dusty tunnel environments. Background Technology

[0002] Tunnel construction environments typically present harsh conditions such as insufficient lighting and pervasive dust, leading to severe degradation in image acquisition quality and blurred target features. This poses a significant challenge to the environmental perception and target recognition capabilities of construction robots. Construction robots must accurately identify construction equipment, procedures, and personnel within the tunnel in real time to ensure operational safety and efficiency.

[0003] Currently, mainstream target detection models (such as the YOLO series) perform excellently in clear environments, but they have the following technical shortcomings in dusty tunnel environments:

[0004] Insufficient environmental robustness: The general model is not optimized for complex interferences such as dim lighting and high dust levels. Under conditions of image quality degradation, the feature extraction capability decreases, and the detection accuracy and robustness are significantly reduced.

[0005] High model complexity: In pursuit of high accuracy, existing models typically have high computational complexity and a large number of parameters, making it difficult to achieve real-time inference on embedded devices mounted on construction robots.

[0006] Lightweight design and accuracy are difficult to balance: Simply making the model lightweight often leads to a decrease in detection accuracy, especially in the recognition of small targets and complex geological features.

[0007] Therefore, there is an urgent need for a lightweight target detection method that can maintain high detection accuracy, adapt to the dusty environment of tunnels, and be suitable for real-time operation on embedded platforms. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a lightweight target detection method and system for construction robots in dusty tunnel environments. This method can improve detection accuracy and robustness in dim, high-dust environments while reducing model complexity.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a lightweight target detection method for a construction robot in a dusty tunnel environment, comprising the following steps:

[0011] Acquire images of the tunnel construction environment and perform preprocessing;

[0012] The preprocessed image is input into a lightweight object detection network to extract features and perform object recognition. The lightweight object detection network includes a backbone network, a neck network, and a detection head.

[0013] The backbone network is equipped with a dynamic asymmetric perception module, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features.

[0014] The neck network is equipped with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features;

[0015] The detection head is equipped with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features and to perform target recognition based on the illumination-compensated features.

[0016] As an alternative implementation, the dynamic asymmetric sensing module includes multiple parallel depthwise separable convolutional branches, with branch convolutional kernels including square convolutional kernels, dilated convolutional kernels, horizontal strip convolutional kernels, and vertical strip convolutional kernels, and the outputs of each branch are weighted and aggregated through dynamic weights.

[0017] As an alternative implementation method, the generation of dynamic weights includes:

[0018] Global average pooling is performed on the input feature map in both the height and width directions to generate weight vectors containing location information. These weight vectors are then input into a multilayer perceptron to generate attention weights for each convolutional branch.

[0019] As an alternative implementation method, deep features and shallow features are cross-fused to obtain fused features, specifically:

[0020] Deep features are concatenated with shallow features, and joint channel attention weights are calculated.

[0021] The weights are split into two parts, one corresponding to deep features and the other to shallow features.

[0022] The two features are fused through a cross-complementary approach: deep features are added to weighted shallow features, and shallow features are added to weighted deep features.

[0023] As an alternative implementation method, adaptive illumination compensation is performed on the fused features, specifically as follows:

[0024] Global average pooling and global max pooling are performed on the fused feature maps respectively, and the pooling results are concatenated into a dual-path feature vector.

[0025] The dual-path feature vectors are input into the multilayer perceptron to generate basic normalized weights;

[0026] Learnable gain coefficients are introduced, and adaptive scaling coefficients are generated through an exponential function to enhance the fused feature map channel by channel.

[0027] As an alternative implementation, the lightweight object detection network is based on the YOLO11n model and is constructed by replacing its original C3k2 module with a dynamic asymmetric sensing module, embedding a cross-complementary fusion module at the lateral connection of the neck network, and connecting a multi-pooling adaptive exposure module in series in front of the detection head.

[0028] Secondly, the present invention provides a lightweight target detection system for a construction robot in a dusty tunnel environment, comprising:

[0029] The data acquisition and preprocessing module is configured to acquire and preprocess images of the tunnel construction environment.

[0030] The target detection module is configured to: input the preprocessed image into a lightweight target detection network to extract features and perform target recognition, wherein the lightweight target detection network includes a backbone network, a neck network, and a detection head;

[0031] The asymmetric perception module is configured such that: the backbone network is provided with a dynamic asymmetric perception module, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features;

[0032] The cross-complementary fusion module is configured such that: the neck network is provided with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features;

[0033] The multi-pooling adaptive exposure module is configured such that: the detection head is provided with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features, and to perform target recognition based on the illumination-compensated features.

[0034] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0035] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0036] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] This invention proposes a lightweight target detection method for construction robots in dusty tunnel environments. Regarding detection accuracy, by introducing a dynamic asymmetric perception module (C3k2-DNPM), a cross-complementary fusion module (CCFM), and a dual-pooling adaptive exposure module (DAEM), the method effectively improves the model's feature extraction and adaptability in dimly lit, high-dust tunnel environments. Compared to the original YOLO11n model, the average accuracy is improved by 3.4%; the F1-score increases from 0.949 to 0.9523, significantly enhancing the robustness of construction equipment recognition under occlusion and interference.

[0039] Regarding model lightweighting, optimizations such as depthwise separable convolution and dynamic weight sharing were implemented to reduce the number of model parameters and model size while maintaining or even improving accuracy, making it more suitable for the computing power limitations of embedded platforms for construction robots. After actual deployment on a tracked inspection robot on a Raspberry Pi 5, the mobile inference speed reached 9.1 FPS, meeting real-time detection requirements. Even in complex tunnel environments, it maintained an average recognition accuracy of 85.4%, verifying the engineering feasibility of this method for efficient and stable operation on edge devices.

[0040] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0042] Figure 1 This is a schematic diagram of the cross-complementary fusion module (CCFM) of the present invention;

[0043] Figure 2 This is a schematic diagram of the structure of the dual pooling adaptive exposure module (DAEM) of the present invention;

[0044] Figure 3 This is a schematic diagram of the lightweight target detection network for the tunnel dust environment construction robot of the present invention;

[0045] Figure 4 This is a schematic diagram of the structure of the dynamic asymmetric sensing module (C3k2-DNPM) of the present invention;

[0046] Figure 5 This is a schematic diagram of the structure of the dynamic asymmetric depthwise separable convolution of the present invention. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0050] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0051] Example 1

[0052] like Figures 1 to 5 As shown, this embodiment provides a lightweight target detection method for construction robots in dusty tunnel environments, including the following steps:

[0053] Acquire images of the tunnel construction environment and perform preprocessing;

[0054] The preprocessed image is input into a lightweight object detection network to extract features and perform object recognition. The lightweight object detection network includes a backbone network, a neck network, and a detection head.

[0055] The backbone network is equipped with a dynamic asymmetric perception module, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features.

[0056] The neck network is equipped with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features;

[0057] The detection head is equipped with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features and to perform target recognition based on the illumination-compensated features.

[0058] The specific solution of the present invention is as follows:

[0059] like Figure 3As shown, the overall network architecture of this invention consists of four parts: an input terminal, a backbone network, a neck network, and a detection head.

[0060] Input: Receives tunnel environment images (default resolution (640 × 640)) and performs preprocessing.

[0061] Backbone: Responsible for feature extraction. All existing C3k2 modules are replaced with C3k2-DNPM modules. This layer contains 5 layers, extracting features at different resolutions (P1-P5). The deeper features (P3-P5) are rich in semantic information and exhibit strong robustness to changes in illumination.

[0062] The neck network employs a PANet (Path Aggregation Network) structure. A CCFM module is deployed at the lateral connections between the upsampled and downsampled features. This module is responsible for "cross-complementary" fusion of deep semantic features (deep features) extracted from the backbone with shallow detail features (shallow features).

[0063] Head Network: The DAEM module is cascaded before the final Detect head. This module performs final refinement and enhancement on the fused features, and then outputs the target's class probability and bounding box coordinates.

[0064] The overall feature extraction data process of this invention is as follows: Figure 3 As shown:

[0065] The feature extraction data stream begins with a raw image. The raw image is input into the backbone network, undergoes downsampling through a series of convolutional modules, and outputs a preliminary shallow feature map. Subsequently, the data enters multiple cascaded C3k2_DNPM modules and convolutions. For each DNPM module, the input feature map is the feature map output from the previous convolutional layer. Internally, the module utilizes multi-scale receptive fields and coordinate attention for processing, resulting in a feature map with the same size but richer spatial location information. The size of the feature map can be changed for each convolution, and interpolation convolutions maintain the three final output feature maps at large, medium, and small sizes. The intermediate output feature maps of the backbone network are P3 and P4. At the end of the backbone network, the feature map output from the last C3k2_DNPM module passes through SPPF (Fast Spatial Pyramid Pooling) and C2PSA modules, outputting a deep semantic feature (P5) after global receptive field expansion and spatial attention enhancement. The backbone network ultimately outputs three feature maps (P3, P4, P5) at different scales to the neck network, representing shallow (large), medium (medium), and deep (small) features, respectively. The feature fusion stage of the neck network is responsible for fusing the three inputs P3, P4, and P5, including both top-down and bottom-up paths. The inputs to both paths are... (Deep features) and (Shallow features), but this varies depending on the processing stage. and Each represents a different characteristic, as detailed below:

[0066] Top-down approach: First, the P5 features are upsampled as the basis for this stage. The P4 characteristics transmitted from the backbone network serve as the basis for this stage. This stage and The input is processed by the CCFM module, which calculates channel attention and then fuses the two, outputting the fused P4-level features. These P4-level features are then refined using a C3k2_DNPM module. Next, the data is upsampled again. The amplified features from the upsampled data are used as... The P3 characteristics transmitted from the backbone network serve as The two are then fused to output the fused P3-level features. Finally, the fused P3-level features are passed through the C3k2_DNPM module, which outputs a high-resolution feature map for detecting small targets, serving as the first final output. Bottom-up path: The output P3-level features are downsampled through convolution (Conv) to serve as... The P4 level features generated in the top-down path are used as The two are then fused to output a secondary fused mesoscale feature map. This output is then processed by the C3k2_DNPM module to produce a mesoscale feature map for detecting mid-range targets, serving as the second final output for the neck area. Next, the data is further downsampled through convolution (Conv) to obtain... The deep features starting from the top-down path (i.e., P5 features after upsampling) are used as... The input is fused into the CCFM module, and the output is the large-scale feature map after secondary fusion. Finally, the large-scale feature map after secondary fusion passes through the C3k2_DNPM module, which outputs a low-resolution feature map for detecting large targets, serving as the third final output of the neck network. The signal then flows to the head network, predicting the output of the three feature maps received from the neck network (i.e., the first, second, and third final outputs). The DAEM detection head network has three DAEM detection heads, corresponding to the three scales (small, medium, and large) of the neck output. For each DAEM module, the input is the feature map of the corresponding scale from the neck network. ( It is also a characteristic of a certain stage, different stages (Representing different features). Internally, the module first extracts features through deep convolution, then combines average pooling and max pooling to generate dual attention weights, outputting a feature map that has been highly enhanced by attention. Final mapping: The enhanced feature map undergoes final... The convolutional layer outputs the final prediction tensor, containing the bounding box coordinates for each grid cell. Confidence level and class probability.

[0067] The specific module settings of this invention are as follows:

[0068] (a) Backbone network: Dynamic asymmetric sensing module (C3k2-DNPM).

[0069] To address the unique technical challenge of detecting targets at extremely polarized scales in tunnel construction scenarios—namely, the need to simultaneously detect tiny targets with very small pixel proportions (such as explosives and anchor bolts) and giant machinery occupying a large area of ​​the image (such as wet spraying trucks and concrete mixer trucks), as well as linear structures with extreme aspect ratios (such as steel arches)—existing technologies using single-scale or conventional multi-scale convolutional kernels cannot handle both simultaneously. To solve the problem that traditional 3×3 square convolutional kernels are insufficient for effectively extracting features of long, narrow, and large equipment within tunnels, this invention designs a dynamic asymmetric sensing module (C3k2-DNPM), as follows:

[0070] 1. Channel grouping and dynamic sensing:

[0071] Input feature map (in For the number of channels, (For height and width), the features are first segmented into and , It is divided into two feature channels. Entering the DynamicNon-symmetricDWConv2d convolutional unit, this unit contains Parallel depthwise separable convolutional branches:

[0072] Branch 1 Shaped convolution with an expansion rate r=1 captures common local features, such as small-sized explosives;

[0073] Branch 2: Dilated convolution with a dilation rate r=2 is used for large machinery such as wet spray trucks and concrete mixer trucks. Dilated convolution is introduced to expand the receptive field.

[0074] Dilated convolution output formula: For the input entering the dilated convolution... and convolution kernel In position Output for:

[0075] ;

[0076] in It is the set of coordinates of the convolution kernel. The expansion rate is denoted as .

[0077] Experience the Wild ( RF Calculate: the receptive field size of a single-layer dilated convolution. RF for:

[0078] ;

[0079] when At that time, feel the wild:

[0080] ;

[0081] Branch 3: Horizontal strip convolution is specifically designed to capture horizontally laid steel arch frames, etc.

[0082] Branch 4: Vertical strip convolution is specifically designed to capture construction equipment such as anchor bolts.

[0083] 2. Dynamic weight generation:

[0084] for Perform feature aggregation operations on both the X and Y axes:

[0085] X-axis feature aggregation (preserving height location information):

[0086] For the One channel, with a height of [missing information]. Output The calculation formula is:

[0087] ;

[0088] This operation iterates through the width direction, compressing the positional information to the height direction, and outputs a feature map of size [size missing]. .

[0089] Y-axis feature aggregation (preserving width position information):

[0090] For the One channel, width is Output The calculation formula is:

[0091] ;

[0092] This operation iterates through the height direction, compressing the positional information to the width direction, and outputs a feature map of size [size missing]. .

[0093] 3. Feature aggregation:

[0094] First, generate the two feature vectors. and A concatenation operation is performed in the spatial dimension to form an aggregated feature map that includes the global view.

[0095] Subsequently, using Convolution Transform Function (Conv) Channel compression and transformation are performed on the aggregated feature map.

[0096] The final four sets of weights are generated using the Softmax normalization function. ,in( )

[0097] Finally, the generated attention weights are applied to the input feature map via element-wise multiplication. This mechanism acts as a feature selector and adaptively “reweights” the feature responses, automatically selecting the more needed feature extraction branches to highlight key targets while suppressing background noise caused by dust and low light.

[0098] Using the weights containing location information generated in step one The outputs of the four branches are dynamically weighted and aggregated:

[0099] ;

[0100] (ii) Neck network: Cross-complementary fusion module (CCFM).

[0101] To address the blurring of target edges caused by tunnel dust obscuring the target, the Cross-Complementary Fusion Module (CCFM) utilizes the complementarity between heterogeneous feature layers for repair.

[0102] 1. Joint attention modeling:

[0103] In deep learning (especially YOLO or multimodal networks), we often need to fuse two feature layers. In YOLO meshes, the Concat module is often used to concatenate two input feature maps to extract features. After replacing Concat with the Cross-Complementary Fusion (CCFM) module... Figure 1 The network structure diagram shown below assumes that the two feature maps to be fused in the input CCFM module are... (Deep features) and (Shallow features). First, the two inputs are concatenated (Concat) so that the network can see information from both inputs simultaneously. Then, the concatenated feature map is subjected to global average pooling. Spatial information compressed into Only channel information is retained. Next, channel attention is calculated using two Pw-Conv (Pointwise Convolution) methods. The Conv module uses two consecutive layers. Convolution. The first layer is usually used for dimensionality reduction and information compression, and the second layer is used for dimensionality increase to restore the number of channels, thereby identifying the feature channels containing key information. Then, attention weights are generated through the sigmoid normalization module.

[0104] The formula for calculating the joint channel attention weight W is as follows:

[0105] ;

[0106] in for The function. The weights W are then decomposed into and So that it can be restarted.

[0107] 2. Cross-Complementary Fusion:

[0108] This invention innovatively employs a "cross-complementary" approach to fuse features, as shown in the following formula:

[0109] ;

[0110] The left-hand side of the formula represents the use of The high-response regions of (shallow detail features) are added as "residuals" to (Deep semantic features) are used to recover edge information lost due to downsampling; the right term utilizes Semantic information filling Voids caused by dust interference.

[0111] (iii) Head Network: Dual Pooling Adaptive Exposure Module (DAEM).

[0112] To address the extremely unstable lighting conditions in tunnel construction environments—specifically, the simultaneous presence of extreme darkness (causing small target features to be obscured) and direct headlight exposure from vehicles (causing localized overexposure)—this invention designs a Dual-pooling Adaptive Exposure Module (DAEM). This module overcomes the limitations of traditional attention mechanisms that rely solely on average features and fixed gain. Through dual-path extreme value sensing and a learnable dynamic exposure mechanism, it achieves feature-level adaptive lighting compensation. Specifically, the DAEM module is as follows:

[0113] 1. Dual-Pooling Perception:

[0114] As a plug-and-play module inserted before the detection head, let the input feature map entering the DAEM module be... To accurately capture the high-frequency reflective features of small targets such as explosives and anchors in dark backgrounds while preserving global background illumination information, this module first passes through a depthwise separable convolutional CMSS module, which is a highly efficient feature extraction block.

[0115] For the CMSS module: First, spatial features are extracted through a deep convolution, then through a GELU and normalization layer, followed by an internal residual connection to prevent deep network degradation, and finally through a... Pointwise convolution integrates information to extract richer contextual information. Then, two pooling operations are performed in parallel:

[0116] Global average pooling (AvgPool): Extracts the overall illumination baseline of the feature map.

[0117] Global Max Pooling (MaxPool): Extracts extreme response points (i.e., highlighted feature points) from the feature map to prevent weak target signals from being diluted by background noise during the averaging process.

[0118] The two pooling results are concatenated along the channel dimension to generate a feature vector containing rich information about illumination distribution. :

[0119] ;

[0120] 2. Channel correlation modeling:

[0121] Will The input is a multilayer perceptron (MLP), which captures the nonlinear dependencies between channels through dimensionality reduction and expansion operations, and generates basic normalized weights via a sigmoid function.

[0122]

[0123] 3. Dynamic Exposure Adaptation:

[0124] To address the challenge of deterministic gain failing to handle alternating periods of direct headlight beams and extreme dimness, this invention introduces a learnable gain coefficient. Construct an adaptive exponential activation function.

[0125] Calculation formula:

[0126]

[0127]

[0128] in, These are parameters that can be automatically updated during network training; initialized to 1.0. This indicates element-wise multiplication.

[0129] In extremely dark areas: the network increases through training. value This allows the scale to grow exponentially, thereby nonlinearly amplifying weak target feature signals and significantly improving the recall rate of small targets in low light.

[0130] In areas directly illuminated by headlights: the network automatically reduces. value This makes the Scale approach 1 or a smaller gain, thereby suppressing overexposure and preventing the loss of target features caused by strong light.

[0131] pass Through dynamic adjustment, the module achieves an "auto exposure" effect similar to the human eye or camera, ensuring the model's robustness under complex lighting conditions.

[0132] (iv) Model training and configuration.

[0133] To verify the effectiveness of this solution under extreme conditions, the present invention uses a high-performance computing platform for model training, with the specific configuration as follows:

[0134] Hardware environment: NVIDIA RTX 5090 GPU.

[0135] Gradient calculation of operators; implementation of low-level operator optimization in conjunction with CUDA 12.6 computation library.

[0136] Software environment: Based on Python 3.13 and the PyTorch deep learning framework.

[0137] Hyperparameter settings:

[0138] Optimizer: SGD (Stochastic Gradient Descent) was selected, and the momentum was set to 0.937 to achieve stable optimization on the non-convex loss surface;

[0139] Learning rate: The initial learning rate is set to Cosine annealing is used for attenuation.

[0140] Iteration rounds: set Because the tunnel environment samples are noisy, and dynamic convolution (DIMB) includes additional attention parameters, longer training cycles are required to ensure that the model converges sufficiently under complex distributions;

[0141] Loss function: CIoU Loss is used as the bounding box regression loss, combined with DFL (DistributionFocal Loss) to optimize the localization accuracy of occluded targets.

[0142] Example 2

[0143] This embodiment provides a lightweight target detection system for construction robots in dusty tunnel environments, including:

[0144] The data acquisition and preprocessing module is configured to acquire and preprocess images of the tunnel construction environment.

[0145] The target detection module is configured to: input the preprocessed image into a lightweight target detection network to extract features and perform target recognition, wherein the lightweight target detection network includes a backbone network, a neck network, and a detection head;

[0146] The asymmetric perception module is configured such that: the backbone network is provided with a dynamic asymmetric perception module, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features;

[0147] The cross-complementary fusion module is configured such that: the neck network is provided with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features;

[0148] The multi-pooling adaptive exposure module is configured such that: the detection head is provided with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features, and to perform target recognition based on the illumination-compensated features.

[0149] It should be noted that the above modules correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above modules and their corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules can be executed in a computer system as part of the system.

[0150] In further embodiments, the following is also provided:

[0151] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0152] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0153] A computer-readable storage medium for storing computer instructions that, when executed by a processor, perform the method of Embodiment 1.

[0154] The method in Example 1 can be directly executed by a hardware processor, or it can be executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0155] A computer program product includes a computer program that, when executed by a processor, implements the method in Embodiment 1.

[0156] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0157] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0158] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0159] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0160] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A lightweight target detection method for construction robots in dusty tunnel environments, characterized in that, Includes the following steps: Acquire images of the tunnel construction environment and perform preprocessing; The preprocessed image is input into a lightweight object detection network to extract features and perform object recognition. The lightweight object detection network includes a backbone network, a neck network, and a detection head. The backbone network is equipped with a dynamic asymmetric perception module, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features. The dynamic asymmetric sensing module includes multiple parallel depthwise separable convolutional branches. The branch convolutional kernels include square convolutional kernels, dilated convolutional kernels, horizontal strip convolutional kernels, and vertical strip convolutional kernels. The outputs of each branch are weighted and aggregated through dynamic weights. The neck network is equipped with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features; The detection head is equipped with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features and to perform target recognition based on the illumination-compensated features. The multi-pooling adaptive exposure module includes a depthwise separable convolutional CMSS module. The fused feature map is input into the CMSS module, where spatial features are extracted via depthwise convolution. Then, it passes through GELU and a normalization layer, followed by an internal residual connection to prevent degradation of the deep network. Finally, it passes through a... Pointwise convolution integrates information and extracts richer contextual information; Global average pooling and global max pooling are performed on the feature maps output by the CMSS module, and the pooling results are concatenated into a dual-path feature vector. The dual-path feature vector is then input into a multilayer perceptron to generate basic normalized weights. A learnable gain coefficient is introduced, and an adaptive scaling coefficient is generated through an exponential function to enhance the fused feature map channel by channel.

2. The lightweight target detection method for tunnel dusty environment construction robots as described in claim 1, characterized in that, The generation of dynamic weights includes: Global average pooling is performed on the input feature map in both the height and width directions to generate weight vectors containing location information. These weight vectors are then input into a multilayer perceptron to generate attention weights for each convolutional branch.

3. The lightweight target detection method for tunnel dusty environment construction robots as described in claim 1, characterized in that, The deep and shallow features are cross-fused to obtain the fused features, specifically: Deep features are concatenated with shallow features, and joint channel attention weights are calculated. The weights are split into two parts, one corresponding to deep features and the other to shallow features. The two features are fused through a cross-complementary approach: deep features are added to weighted shallow features, and shallow features are added to weighted deep features.

4. The lightweight target detection method for tunnel dusty environment construction robots as described in claim 1, characterized in that, The lightweight object detection network is based on the YOLO11n model. It is constructed by replacing the original C3k2 module with a dynamic asymmetric sensing module, embedding a cross-complementary fusion module at the lateral connection of the neck network, and connecting a multi-pooling adaptive exposure module in series in front of the detection head.

5. A lightweight target detection system for construction robots in dusty tunnel environments, characterized in that: include: The data acquisition and preprocessing module is configured to acquire and preprocess images of the tunnel construction environment. The target detection module is configured to: input the preprocessed image into a lightweight target detection network to extract features and perform target recognition, wherein the lightweight target detection network includes a backbone network, a neck network, and a detection head; The asymmetric perception module is configured such that: a dynamic asymmetric perception module is set in the backbone network, which is used to extract image features through asymmetric convolution kernels to obtain deep features and shallow features; the dynamic asymmetric perception module includes multiple parallel depthwise separable convolution branches, and the branch convolution kernels include square convolution kernels, dilated convolution kernels, horizontal strip convolution kernels and vertical strip convolution kernels, and the outputs of each branch are weighted and aggregated through dynamic weights; The cross-complementary fusion module is configured such that: the neck network is provided with a cross-complementary fusion module, which is used to cross-fuse deep features and shallow features to obtain fused features; The multi-pooling adaptive exposure module is configured such that: the detection head is provided with a multi-pooling adaptive exposure module, which is used to perform adaptive illumination compensation on the fused features and to perform target recognition based on the illumination-compensated features; The multi-pooling adaptive exposure module includes a depthwise separable convolutional CMSS module. The fused feature map is input into the CMSS module, where spatial features are extracted via depthwise convolution. Then, it passes through GELU and a normalization layer, followed by an internal residual connection to prevent degradation of the deep network. Finally, it passes through a... Pointwise convolution integrates information and extracts richer contextual information; Global average pooling and global max pooling are performed on the feature maps output by the CMSS module, and the pooling results are concatenated into a dual-path feature vector. The dual-path feature vector is then input into a multilayer perceptron to generate basic normalized weights. A learnable gain coefficient is introduced, and an adaptive scaling coefficient is generated through an exponential function to enhance the fused feature map channel by channel.

6. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-4.

8. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-4.

Images