A method for detecting drivable areas adapted to mining scenarios

Abstract

This invention discloses a drivable area detection method adapted to mining scenes, belonging to the field of image detection and segmentation. The method processes the input image by constructing a semantic segmentation detection model, including a lightweight encoder, a multi-scale feature fusion module, and a multi-attention decoder. The encoder uses an optimized MobileNetV3 backbone network, optimized through dynamic stride depth-separable convolution design and mining scene channel pruning; the decoder integrates a dual-dimensional enhanced attention module (M-DAM) and an SE channel attention module, using dual-dimensional weighted suppression of interfering features and enhancement of effective features. Experiments show that this method achieves a detection rate of 22.45 FPS on a mining dataset, with an average intersection-over-union ratio of 0.7008, balancing lightweight and low computational requirements with high accuracy and strong real-time performance. It effectively solves the fragmentation and boundary blurring problems of traditional models, adapting to the transportation safety requirements of unstructured road scenes in mines.

Description

Technical Field

[0001] This invention relates to the field of image detection and segmentation, and in particular to a method for detecting drivable areas adapted to mining scenes. Background Technology

[0002] Anomalies in drivable areas within mining environments can be caused by mining disturbances, extreme weather, or dynamic topographical changes, representing a current source of safety hazards that cannot be effectively avoided in mine transportation. Due to the complex terrain and frequent environmental changes in most mines in my country, the safety of transport vehicles faces serious threats. Therefore, to reduce mining transportation risks and ensure traffic safety, this research focuses on drivable area detection technology for mining environments. It utilizes intelligent algorithm models to accurately determine the status of these areas, and validates the detection model through image data acquisition and analysis. To improve the accuracy and real-time nature of the determination results, data acquisition and analysis must be both accurate and rapid.

[0003] Existing drivable area detection technologies are mostly based on general semantic segmentation models. Although lightweight technologies such as depthwise separable convolution and attention mechanisms have been adopted for real-time detection of drivable areas, they have not been optimized for the mining environment and still suffer from problems such as an imbalance between accuracy and real-time performance, incompatibility with the deployment requirements of low-computing-power equipment, and insufficient robustness. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a lightweight, low-computing-power, high-precision and real-time drivable area detection method that is more adapted to mining scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for detecting drivable areas adapted to mining scenarios, the key of which lies in...

[0007] By constructing a semantic segmentation detection model, the input image is identified, and the output image marks the drivable area;

[0008] The semantic segmentation detection model includes a lightweight encoder, a multi-scale feature fusion module, and a multi-attention decoder;

[0009] The input image is processed by a lightweight encoder and a multi-scale feature fusion module to extract fused features from the image processed by the lightweight encoder; the multi-attention decoder performs feature recognition on the feature map output by the multi-scale feature fusion module, and finally outputs an image marking the drivable area;

[0010] The lightweight encoder employs an optimized MobileNetV3 backbone network, with optimizations including dynamic stride depth separable convolution design and mining scene channel pruning; the multi-attention decoder includes a dual-dimensional enhanced attention module M-DAM and an SE channel attention module.

[0011] Preferably, the dynamic stride depth separable convolution design optimizes the original fixed stride into a region-adaptive dynamic stride, dividing the image edge and center regions by pixel coordinates and setting the stride differently.

[0012] The mine scene channel pruning method adds a mine feature corresponding threshold screening mechanism on the basis of the original NAS channel pruning. It calculates the response value R of each channel to the drivable area features, dynamically retains channels with R≥T, where T is the dynamic threshold, and prunes redundant channels corresponding to interference channels.

[0013] Preferably, the NAM spatial-channel dual attention module uses spatial-channel dual-dimensional feature weighting;

[0014] In terms of spatial dimension: by using the batch normalization scaling factor λ, low weights are assigned to locally clustered interference regions in the image, thus weakening their feature responses;

[0015] Channel dimension: By using the batch normalization scaling factor γ, high weights are assigned to channels (such as road surface texture and alleyway outline) in the drivable area to enhance effective features;

[0016] The calculation of the channel and spatial attention mechanism submodule is shown below:

[0017]

[0018] in This represents the output feature, assuming the feature has c channels and weights of .

[0019]

[0020] In the formula This represents the output feature, with weights of .

[0021]

[0022] In the formula Indicates input, It is the output. Indicates network weights, This represents the forward prediction function of the model. (•) is the loss function. (•)yes Norm penalty function; It is balance and The penalty items.

[0023] Preferably, the normalized scaling factor is calculated as follows:

[0024]

[0025] In the formula, and They are small batches The mean and standard deviation; and These are trainable affine transformation parameters. This is a constant added to the mini-batch variance to ensure the stability of the values.

[0026] Preferably, the SE channel attention module adds a "squeeze-excitation" channel filtering mechanism after the NAM module to perform secondary optimization on the global features;

[0027] The squeezing phase involves compressing the feature dimensions through global average pooling, focusing on the global semantic information of the drivable region.

[0028] Incentive phase: The channel weights are dynamically learned through two fully connected layers to further strengthen channels that are strongly correlated with the drivable area and suppress channels that only respond to interference.

[0029] The beneficial effects of adopting the above technical solution are as follows:

[0030] In this invention, the encoder part adopts the MobileNetV3 network and performs dynamic stride depth separable convolution design and adaptive optimization of channel pruning in mining scenarios. While reducing the number of parameters to less than 6M, the detection accuracy of small targets is maintained, which meets the requirements of lightweight, low computing power and strong real-time performance of detection methods in mining scenarios.

[0031] This invention constructs a multi-attention collaborative mechanism in the decoder section and optimizes the feature extraction logic in a targeted manner for the characteristics of interference objects in mining scenarios, thereby reducing the false judgment rate and improving the accuracy of drivable area detection. Attached Figure Description

[0032] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0033] Figure 1 This is a schematic diagram of the network structure of the NSM3-DeepLabV3+ model proposed in this invention;

[0034] Figure 2 This is a visualization of the hard activation function in this invention;

[0035] Figure 3 This is a schematic diagram of the MobileNetV3 backbone network in this invention;

[0036] Figure 4 This is a schematic diagram of the channel attention mechanism in the dual-dimensional enhanced attention module M-DAM of this invention;

[0037] Figure 5 This is a schematic diagram of the spatial attention mechanism in the dual-dimensional enhanced attention module M-DAM of this invention;

[0038] Figure 6 This is a schematic diagram of the network structure of the NACM3-DeepLabV3+ model proposed in this embodiment of the invention;

[0039] Figure 7 This is a schematic diagram of the ACmix attention mechanism proposed in this embodiment of the invention;

[0040] Figure 8 This is a schematic diagram of the C2f-SCConv attention mechanism in an embodiment of the present invention;

[0041] Figure 9 This is a visualization result of the ablation experiment in the embodiments of the present invention;

[0042] Figure 10 This is a visualization result of the comparative experiment in the embodiments of the present invention. Detailed Implementation

[0043] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention proposes a method for detecting drivable areas adapted to mining scenes. By constructing a semantic segmentation model, the input mining scene image is processed, and the image information of the segmented and marked drivable areas is output.

[0045] Regarding the semantic segmentation model, this invention proposes a lightweight semantic segmentation model, NSM3-DeepLabV3+, such as... Figure 1 It adopts a unified architecture design of lightweight encoder, multi-scale feature fusion module ASPP and multi-attention decoder.

[0046] The NSM3-DeepLabV3+ model uses an optimized MobileNetV3 network in the backbone network.

[0047] The introduction of hard-Swish / hard-Sigmoid activation functions and SE attention mechanism into MobileNetV3 has achieved initial performance improvements. Hard activation functions can improve the non-linear expressiveness of the model while reducing computational complexity, thereby improving overall computational efficiency.

[0048] The core idea of ​​hard-Swish is to add a constant offset to the input and then make corrections; while preserving some of the nonlinear characteristics of ReLU, it achieves a smoother transition, thereby reducing the possibility of gradient vanishing. The specific formula is as follows:

[0049]

[0050] The hard-Sigmoid model achieves a smooth activation effect similar to the Sigmoid model through ReLU6, which helps the model to quickly compute the output during the inference phase.

[0051]

[0052] like Figure 2 As shown, hard-Swish and hard-Sigmoid combine low computational overhead with non-linear expressive power, reducing the risk of gradient vanishing through smooth transitions and improving model efficiency and accuracy.

[0053] like Figure 3 The SE attention mechanism dynamically weights feature channels through a process of global average pooling, fully connected dimensionality reduction, ReLU activation, fully connected dimensionality increase, and weighting operations, thereby enhancing the ability to effectively represent features.

[0054] During the research and development process, the researchers found that MobileNetV3 has the advantage of being lightweight, but its general architecture is not adapted to the characteristics of "small targets dense and complex interference" in mining scenarios. Therefore, this invention optimizes the above-mentioned introduction of hard activation functions and SE attention mechanisms, including dynamic stride depth separable convolution design and channel pruning in mining scenarios.

[0055] Dynamic stride depth-separable convolution: MobileNetV3 uses fixed stride convolution, which leads to the loss of features of small targets such as narrow mine tunnels and small drivable areas. The fixed stride is optimized to a region-adaptive dynamic stride. The image edge region and center region are divided by pixel coordinates, and the stride is different for different regions. This can improve the accuracy of small target detection in mines without increasing parameters, thus balancing lightweight and accuracy.

[0056] Mine scene channel pruning: The NAS channel pruning of MobileNetV3 is designed for general scenarios and does not filter out mine interference, such as redundant channels corresponding to interference like gravel and vegetation. Based on the original channel pruning of MobileNetV3, a mine feature response threshold screening mechanism is added, that is, calculate the response value R of each channel to the "drivable area feature", dynamically set the threshold T, and only retain the channels with R≥T, pruning the interference channels with R<T. Through channel pruning, the parameter quantity is further reduced, and at the same time, the response of the drivable area feature is strengthened, laying a foundation for subsequent attention anti-interference optimization.

[0057] The multi-attention decoder includes a mine-adapted two-dimensional enhanced attention module M-DAM and an optimized SE channel attention module.

[0058] The two-dimensional enhanced attention module M-DAM aims at the characteristics of "local aggregation and similar grayscale" of mine interference such as gravel and dust, and precisely suppresses the interference feature and strengthens the drivable area feature through spatial-channel two-dimensional feature weighting.

[0059] Among them, for the channel dimension: based on the batch normalization adaptive scaling parameter γ, introduce the mine feature adaptation coefficient η, and assign high weights to the channels that respond to the drivable area (such as road surface texture and roadway contour) to strengthen the effective features;

[0060] For the spatial dimension: use the batch normalization scaling parameter λ, combined with the mine interference density factor ζ, assign low weights to the areas with dense interference, and at the same time, through spatial neighborhood smoothing processing, weaken the response of invalid features;

[0061] Such as Figure 4-5 , the calculation of the channel and spatial attention mechanism sub-module is as follows:

[0062]

[0063] Among them represents the channel attention output feature. Assuming the feature has c channels, the channel weight is , is the scaling parameter of the i-th channel of the batch normalization layer, is the mine feature adaptation coefficient of the i-th channel (the channels related to asphalt pavement =1.2, the channels related to dirt roads =1.0, the channels related to gravel roads =0.9); is the high-dimensional semantic feature map after the encoder is pruned by the mine scene channel;

[0064]

[0065] In the formula Represents the spatial attention output features, with spatial weights as follows: , The scaling parameter for the i-th spatial location of the batch normalization layer ( ϵ[0.1,1.0]), The density factor of mining disturbances at the i-th spatial location (dense disturbance area) =0.5, road surface area =0.1); This is a detail-enhanced feature map after multi-scale fusion; (·) represents the batch normalization operation for spatial dimensions; G is a 3×3 Gaussian kernel matrix (specifically...). ), used to suppress false activation of isolated interfering substances; Indicates the convolution operation;

[0066] The normalized scaling factor is defined as follows:

[0067]

[0068] In the formula, and They are small batches The mean and standard deviation; and These are trainable affine transformation parameters. This is a constant added to the mini-batch variance to ensure the stability of the values.

[0069] In the formula The input image represents a mining scene. Label the drivable areas. These represent the learnable weight parameters of the entire network. This represents the forward prediction function of the model. (·) represents the semantic segmentation loss function (cross-entropy loss is used for classification tasks). The loss is the boundary alignment loss of the mine road (the error between the predicted boundary and the actual boundary is calculated based on the Hausdorff distance). =0.3 is the loss balance coefficient; g Norm regularization constraint function; =0.01、 =0.005 is the regularization weight coefficient ( Prioritize the constraint of redundant features in the channel dimension). This is the set of scaling parameters for the channel dimensions of the M-DAM module. This is the set of spatial dimension scaling parameters for the M-DAM module.

[0070] The SE channel attention module and the M-DAM module form a mining scene-specific collaborative mechanism of "local optimization-global enhancement": The M-DAM module first performs feature screening for local interference in the mine, and the SE module captures the semantic correlation of long-distance road surfaces in the mine through global average pooling. Then, through two fully connected layers, the channel weights are dynamically learned to perform secondary weighting on the output features of M-DAM, which strengthens the continuity of global features in the drivable area and further suppresses redundant channels that only respond to interference.

[0071] In specific implementation, such as Figure 6 The researchers proposed a performance comparison between the NACM3-DeepLabV3+ model and the NSM3-DeepLabV3+ model. Both models use the same architecture of lightweight encoder + multi-scale feature fusion module (ASPP) + multi-attention decoder. The performance difference is obtained by differentiating the multi-attention decoder. The multi-attention decoder part of the NACM3-DeepLabV3+ model combines M-DAM, ACmix and C2f-SCConv attention mechanisms; the two-dimensional enhanced attention module M-DAM is the same as that in the NSM3-DeepLabV3+ model.

[0072] To address the issue of blurred boundaries between drivable areas and interfering objects in extreme scenarios, the NACM3-DeepLabV3+ model introduces the ACmix hybrid attention module, which combines the advantages of self-attention and convolution to achieve efficient enhancement of feature representation. Specifically, for example... Figure 7 ACmix Hybrid Attention Module The feature map is used as input, and feature interaction is constructed through feature projection and splitting, dual-path processing, and weighted fusion. The structure is as follows: Figure 7 As shown.

[0073] Feature projection and splitting: using 3 The convolutional module maps and segments the input features, decomposing the feature with C channels into... Zhang Zi's characteristics.

[0074] The dual-path processing includes a convolutional branch and a self-attention branch; the convolutional branch involves fusing sub-features with a fully connected layer and a shift operation, simulating the "sliding window" characteristic of convolution, extracting locally correlated features, and outputting N sets of data. feature.

[0075] Self-attention branch: Sub-features respectively assume the roles of Query, Key, and Value. Through multi-head self-attention (MHSA), key regions are dynamically focused to capture long-distance dependencies, and N sets are output. feature.

[0076] Weighted fusion: After concatenating the results of two paths, a learnable scalar is used. and Weighted fusion adaptively balances the fusion ratio of local and global features.

[0077] Final output :

[0078]

[0079] in, It is a self-attention path feature. It is a convolutional branch feature.

[0080] The ACmix hybrid attention module achieves the complementary advantages of self-attention and convolution mechanisms with lightweight computation through a design of shared projection, path-based processing and fusion. It not only strengthens the modeling of long-distance semantic continuity in drivable areas, but also accurately captures local details in scenes with blurred boundaries, effectively enhancing the model's ability to express complex mine features.

[0081] Furthermore, to address the issue of feature redundancy caused by multi-scale disturbances (rubble, vegetation, etc.) in extreme mining scenarios, a C2f-SCConv composite attention module is introduced, such as... Figure 8 Through dual-path optimization of "feature diversion + spatial-channel reconstruction", redundant information is precisely suppressed and effective features are enhanced.

[0082] The C2f-SCConv composite attention module includes CBS, Split, GAM-neck, and SCConv convolution modules. The specific workflow is as follows:

[0083] CBS preprocessing: By combining "convolution (Conv) + batch normalization (BN) + SILU activation", multi-scale features are extracted and the model convergence is accelerated. Nonlinearity is added to adapt to complex mining scenarios.

[0084] Feature splitting (Split): The preprocessed features are split into two paths - the basic information stream directly retains the original details, and the deep feature stream enters the Bottleneck module to extract abstract semantics, so as to take into account the features at different levels.

[0085] Cross-dimensional enhancement (GAM-neck): By connecting the channel and spatial attention sub-modules, it enhances the cross-dimensional interaction of features and retains key dimensional information;

[0086] Spatial-Channel Reconstruction (SCConv): Spatial Reconstruction Units (SRUs) suppress spatial redundancy (such as overlapping gravel areas) through a "separation-reconstruction" approach, while Channel Reconstruction Units (CRUs) filter key channels (such as channels responding to smooth road surfaces) through a "segmentation-transformation-fusion" strategy. The two work together to reduce redundancy and improve feature representation capabilities.

[0087] C2f-SCConv, through its lightweight design with multi-component collaboration, effectively reduces the misclassification rate caused by multi-scale interference without significantly increasing computational load, providing more accurate feature support for subsequent segmentation tasks.

[0088] To verify the differences in the actual performance of the models, the researchers conducted experiments, as follows:

[0089] I. Model Training

[0090] 1) Dataset settings

[0091] This embodiment uses a dataset of 769 mining environment images for annotation. The annotation process involves labeling different target categories, such as sky, vehicles, drivable areas, slopes, obstacles, flags, houses, and pedestrians. The dataset is randomly divided into a training set:test set ratio of 4:1.

[0092] 2) Experimental Environment

[0093] The model proposed in this paper is implemented based on the PyTorch open-source framework, and the environment configuration is shown in Table 1.

[0094] Table 1 Environmental Configuration

[0095]

[0096] 3) Parameter settings

[0097] The experimental parameters in this paper are set as follows: batch sample size: 16 in the frozen phase, 8 in the unfrozen phase, 100 epochs of model iterations, SGD optimizer used for training, initial learning rate: 0.004, minimum learning rate: 0.00004, cosine annealing algorithm is used to dynamically adjust the learning rate, and the internal momentum parameter of the optimizer is set to 0.9.

[0098] 4) Evaluation Indicators

[0099] The model performance was comprehensively evaluated by using the average intersection-to-union ratio (MIoU), average pixel accuracy (mPA), pixel accuracy accuracy, detection rate (FPS), and the number of parameters.

[0100] II. Ablation Experiment

[0101] To evaluate the contribution of the proposed NSM3-DeepLabV3+ network to module performance, ablation experiments were systematically conducted. The experiments focused on analyzing the impact of the improved backbone network on model efficiency, as well as the effects of the two-dimensional enhanced attention module M-DAM, the ACmix attention mechanism, the C2f-SCConv attention mechanism, and the SE attention mechanism on model performance.

[0102] Table 2 Ablation test performance results

[0103] As shown in Table 2, the ablation experiment used DeepLabV3+ (Xception) as the benchmark. By gradually replacing the backbone network and adding an attention mechanism, the impact of each module on the segmentation performance of the drivable area in the mining scene was verified. The core results are as follows:

[0104] Experiment 1 (Lightweight Replacement of MobileNetV3): The backbone network was replaced with MobileNetV3, reducing the number of parameters from 54.711M to 5.815M (a reduction of 89.4%), and increasing the inference speed from 8.96FPS to 20.39FPS (an improvement of 127.6%). The cost was a 4.6% decrease in mIoU, a 3.8% decrease in mPA, and a 0.9% decrease in accuracy, achieving an initial balance between lightweight design and accuracy.

[0105] Experiment 2 (+M-DAM): Based on Experiment 1, the M-DAM mechanism was added, which improved mIoU by 2.52%, mPA by 2.27%, and accuracy by 1.01%, effectively suppressing background interference and improving boundary clarity. The number of parameters remained basically unchanged, and the inference speed was reduced to 16.85 FPS (a decrease of 17.36%), achieving a significant improvement in accuracy at the cost of a small speed.

[0106] Experiment 3 (+ACmix attention): ACmix was added to Experiment 2. Due to feature conflicts, the accuracy decreased slightly, but the inference speed increased to 19.02 FPS (an improvement of 12.9%). The optimization of the computation process offset some of the attention overhead.

[0107] Experiment 4 (+C2f-SCConv Attention): Based on Experiment 3, C2f-SCConv was added, resulting in a slight improvement of mIoU by 0.37% and a further improvement in inference speed to 21.96 FPS (an improvement of 15.46%). Through spatial-channel reconstruction and grouped convolution design, the efficiency of feature fusion and computational complexity were optimized.

[0108] Experiment 5 (+SE Attention): Based on Experiment 2, the SE mechanism was added. The accuracy decreased slightly compared to Experiment 2, but the speed was significantly improved by 33.23%. Compared with Experiment 4, it achieved comparable accuracy and better speed with a simpler module combination, which is suitable for the real-time navigation needs of mines with low computing power.

[0109] like Figure 9 It can be observed that NSM3-DeepLabV3+ segmentation is clear and regular, and performs best in unstructured road scenes.

[0110] III. Comparative Experiment

[0111] To verify the performance of the proposed lightweight model in the unstructured road drivable area segmentation task, a comparative experiment was conducted between the NSM3-DeepLabV3+ model and typical models in the semantic segmentation and detection fields such as UNet, PSPNet, SegFormer, and Hrnet.

[0112] Table 3 Comparison of experimental performance results

[0113] The experimental results are shown in Table 3. The performance comparison analysis of each model in the unstructured road scenario is as follows:

[0114] UNet: It is slightly better than NSM3-DeepLabV3+ in traditional metrics, but the latter has a 47.3% increase in FPS and a 76.7% reduction in parameters, showing significant advantages in speed and lightweight design.

[0115] PSPNet: Regardless of whether ResNet50 or MobileNetV3 backbone is used, fixed pyramid pooling cannot adapt to the dynamic feature requirements of unstructured roads. Its generalization and anti-interference ability are weak, and its accuracy is not as good as the improved model.

[0116] SegFormer-B0: The lightweight Transformer structure has redundant self-attention computations, resulting in slightly slower speed. NSM3-DeepLabV3+ improves FPS by 8.1%.

[0117] HRNet: The multi-resolution parallel architecture has extremely high computational cost and is difficult to implement; the improved model achieves more than 3 times the speed increase and 80% reduction in parameters with a 0.37% difference in mIoU, and the "accuracy-efficiency" balance is more in line with actual needs.

[0118] like Figure 10 UNet (VGG16) has rich semantic information but coarse details, and the semantic segmentation of trucks is incomplete; the PSPNet series either oversimplifies the segmentation, has messy boundaries, or makes incorrect segmentation, and has weak anti-interference ability; SegFormer-B0 and HRNet have fragmented segmentation boundaries, are easily affected by background noise, and have insufficient generalization ability; NSM3-DeepLabV3+ has regular segmentation, strong anti-interference ability, and is more suitable for unstructured road scenes.

[0119] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting drivable areas adapted to mining scenarios, characterized in that, By constructing a semantic segmentation detection model, the input image is identified, and the output image marks the drivable area; The semantic segmentation detection model includes a lightweight encoder, a multi-scale feature fusion module, and a multi-attention decoder; The input image is processed by a lightweight encoder and a multi-scale feature fusion module to extract fused features from the image processed by the lightweight encoder; the multi-attention decoder performs feature recognition on the feature map output by the multi-scale feature fusion module, and finally outputs an image marking the drivable area; The lightweight encoder employs an optimized MobileNetV3 backbone network, with optimizations including dynamic stride depth separable convolution design and mine scene channel pruning. The multi-attention decoder includes a mining-adapted dual-dimensional enhanced attention module M-DAM and an optimized SE channel attention module; The M-DAM dual-dimensional enhanced attention module adapted for mines performs feature selection for localized interference in mines through a two-stage optimization process of "channel dimension feature enhancement + spatial dimension noise suppression". The space Dimension: By using the batch normalization scaling parameter λ and combining it with the mine interference density factor ζ, low weights are assigned to areas with dense interference, while ineffective feature responses are weakened through spatial neighborhood smoothing. Channel dimension: Based on the batch normalized adaptive scaling parameter γ, a mine feature adaptation coefficient η is introduced to assign high weights to channels in the drivable response area, thereby enhancing effective features; The calculation of the channel and spatial attention mechanism submodule is shown below: ； in This represents the channel attention output feature, where c is the number of channels in the feature, and the weights of the channel attention module are... , This represents the scaling parameter corresponding to the i-th channel in the batch normalization layer. Let be the mine feature adaptation coefficient for the i-th channel; This is the high-dimensional semantic feature map of the encoder after pruning through the mine scene channel; BN is the batch normalization operation, which normalizes the input features. ； Represents the spatial attention output features, with spatial weights as follows: , For the scaling parameter of the i-th spatial location in the batch normalization layer, Let be the density factor of the mining disturbance at the i-th spatial location; This is a detail-enhanced feature map after multi-scale fusion; (·) represents the batch normalization operation for spatial dimensions; G is a 3×3 Gaussian kernel matrix, specifically... Used to suppress false activation of isolated interfering substances; Indicates the convolution operation; ； In the formula The input image represents a mining scene. Label the drivable areas. These represent the learnable weight parameters of the entire network. This represents the forward prediction function of the model. (·) represents the semantic segmentation loss function; for classification tasks, cross-entropy loss is used. Loss due to alignment issues with mine road boundaries g is the loss balance coefficient; Norm regularization constraint function; , Here are the regularization weight coefficients, and Prioritize the constraint of redundant features in the channel dimension; This is the set of scaling parameters for the channel dimensions of the M-DAM module. This is the set of spatial dimension scaling parameters for the M-DAM module; The normalized scaling factor is defined as follows: ； In the formula, and They are small batches The mean and standard deviation; and These are trainable affine transformation parameters. This is a constant added to the mini-batch variance to ensure the stability of the values; BN(·): Batch normalization operation, which normalizes the input features; : Mini-batch input data for the batch normalization layer; : Mini-batch output data of the batch normalization layer, i.e., the feature results after BN processing.

2. The method for detecting drivable areas adapted to a mining scenario according to claim 1, characterized in that, The dynamic stride depth-separable convolution design optimizes the original fixed stride into a region-adaptive dynamic stride, dividing the image edge and center regions by pixel coordinates and setting the stride differently. The mine scene channel pruning method adds a mine feature corresponding threshold screening mechanism on the basis of the original NAS channel pruning. It calculates the response value R of each channel to the drivable area features, dynamically retains channels with R≥T, where T is the dynamic threshold, and prunes redundant channels corresponding to interference channels.

3. The method for detecting drivable areas adapted to a mining scenario according to claim 1, characterized in that, The SE channel attention module and the M-DAM module form a "local optimization-global enhancement" collaborative mechanism for mining scenes: the SE module captures the semantic correlation of long-distance road surfaces in the mine through global average pooling, and then dynamically learns the channel weights through two fully connected layers to perform secondary weighting on the output features of M-DAM, thereby strengthening the continuity of global features in the drivable area and further suppressing redundant channels that only respond to interference.

Images