A method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles

Abstract

This invention discloses a method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles in the field of obstacle detection technology. The steps are as follows: First, the YOLOv8 Backbone is used to extract features from the input off-road surface environment containing positive and negative obstacles, thereby obtaining different feature layer information of positive and negative obstacles. The features of the 2nd, 4th, 6th, and 9th layers are denoted as , , , and , respectively. Second, the four feature maps are sequentially fed into the constructed multi-scale fine-grained feature purification branch, multi-scale coarse-grained feature extraction branch, and multi-scale feature fusion strategy. Finally, the feature map after the multi-scale feature fusion strategy is fed into the Head part to obtain the detection results of positive and negative obstacles on the off-road surface. This invention achieves accurate identification of positive and negative obstacles on off-road surfaces, providing key information for subsequent intelligent decision-making and control of unmanned emergency rescue vehicles.

Description

Technical Field

[0001] This invention relates to the field of obstacle detection technology, specifically a method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles. Background Technology

[0002] Emergency rescue vehicles often need to perform rescue missions in off-road environments, such as earthquake rescue, forest fire rescue, and mudslide rescue. The complex and varied terrain of off-road environments poses a threat to the safety of drivers, making the application of unmanned emergency rescue vehicles particularly important. During operation, the perception technology of unmanned emergency rescue vehicles acts like their eyes, enabling them to perceive the surrounding environment in real time and react quickly and accurately to ensure safe driving. However, in off-road environments, the diverse and complex positive and negative obstacles (positive obstacles such as mounds of earth, rocks, and trees on the ground, and negative obstacles such as potholes and ditches) greatly increase the difficulty for unmanned vehicles to perceive their surroundings. Therefore, research on methods for detecting positive and negative obstacles on off-road surfaces by unmanned emergency rescue vehicles is urgently needed.

[0003] Deep learning, with its advantages of automatic feature extraction, excellent generalization performance, high accuracy, and strong robustness, has gained widespread favor in the field of obstacle detection in off-road environments. Deep learning-based detection methods can be divided into single-stage and two-stage methods. While the existing two-stage detection method, Fast-RCNN, which is an improvement on Fast-RCNN, achieves high-precision obstacle detection, its detection speed is relatively slow, making it unsuitable for autonomous vehicles. In contrast, the single-stage method YOLO (You Only Look Once) series strikes a balance between detection accuracy and speed, and therefore has attracted widespread attention in obstacle detection.

[0004] Off-road obstacles are diverse, with significant differences in characteristics between positive and negative obstacles. Positive obstacles (such as mounds, bushes, rocks, and trees) exhibit large scale variations, while negative obstacles (such as pits, trenches, and puddles) are susceptible to visual noise interference due to changes in lighting and water reflection. To address these characteristics, various positive and negative obstacle detection models based on improved YOLO have emerged. To solve the problem of blurred feature information in negative obstacles under complex lighting conditions, the improved YOLOv5 negative obstacle detection model achieves good detection results through path enhancement and dual feature fusion, but it still does not fully consider the multi-scale features of positive obstacles. Furthermore, the improved YOLOv7-tiny model enhances the network's ability to perceive feature maps, enabling simultaneous detection of positive and negative obstacles, but it mainly focuses on positive obstacles with small scale differences and negative obstacles in urban environments, and cannot be applied to positive and negative obstacle detection in off-road environments. Therefore, when designing a positive and negative obstacle detection network for off-road environments, it is essential to thoroughly consider the combined effects of scale differences and noise interference to ensure the accuracy of the detection system. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] Therefore, the purpose of this invention is to provide a method for identifying positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles, so as to achieve accurate identification of positive and negative obstacles on off-road surfaces and provide key information for subsequent intelligent decision-making and control of unmanned emergency rescue vehicles.

[0007] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0008] A method for detecting positive and negative obstacles on off-road surfaces by an unmanned emergency rescue vehicle, comprising the following steps:

[0009] S1. First, the YOLOv8 Backbone is used to extract features from the input off-road environment containing positive and negative obstacles, thereby obtaining different feature layers of positive and negative obstacles. The features at layers 2, 4, 6, and 9 are denoted as follows: , , , ;

[0010] S2. Next, the four feature maps are sequentially fed into the constructed multi-scale fine-grained feature purification branch, multi-scale coarse-grained feature extraction branch, and multi-scale feature fusion strategy. The multi-scale fine-grained feature purification branch is used to reduce noise interference and extract multi-scale local information, the multi-scale coarse-grained feature extraction branch is used to extract multi-scale global information, and the multi-scale feature fusion strategy makes the feature map fed into the Head part contain more target feature information.

[0011] S3. Finally, the feature map processed by the multi-scale feature fusion strategy is fed into the Head part to obtain the positive and negative obstacle detection results on the off-road surface.

[0012] As a preferred embodiment of the method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles described in this invention, the multi-scale fine-grained feature purification branch is composed of multi-receptive field noise filtering feature learning and information interleaving.

[0013] As a preferred embodiment of the off-road obstacle detection method for unmanned emergency rescue vehicles described in this invention, the multi-receptor field noise filtering feature learning consists of a noise filter and multi-receptor field feature learning, and the multi-receptor field noise filtering feature learning performs the following steps:

[0014] Let the input feature map be The output feature map is First, use a convolution kernel size of convolutional layers The number of channels was adjusted, and then the channels of the transformed feature map were separated. The feature map was obtained through equation (1). , , , And using four different sized convolution kernels respectively. , , , Multi-receptive field feature extraction is performed on four feature maps to generate feature maps. , , , ;

[0015] Next, the differential attention mechanism is applied to these feature maps using formula (2) to obtain the feature maps. , , , ;

[0016] (1)

[0017] (2)

[0018] In the formula, Indicates channel segmentation; This represents the differential attention mechanism; Indicates the kernel size as Convolution operations;

[0019] Secondly, for , , , Feature concatenation is performed along the channel dimension, and the order of feature channels is shuffled through a channel shuffling operation before being applied again. Convolution, and the feature map is obtained through equation (3). ;

[0020] (3)

[0021] In the formula, Indicates channel splicing. Indicates mixed washing of channels;

[0022] Finally, the processed feature map is added to the original feature map, which not only helps alleviate the gradient vanishing problem in deep networks, but also promotes cross-layer information interaction. The added feature map is then fed into the MRL to further extract multi-receptive field features. The output feature map is obtained through expression (4). ;

[0023] (4)

[0024] in, This indicates multi-receptor field feature learning.

[0025] As a preferred embodiment of the method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles according to the present invention, the differential attention mechanism performs the following steps:

[0026] Let the input feature map be The output feature map is Then, the difference feature map is generated through equation (5). ;

[0027] (5)

[0028] In the formula, This indicates a bicubic interpolation upsampling operation; This indicates an average pooling downsampling operation; This indicates element-wise subtraction.

[0029] After differential processing, the Sigmoid activation function is applied to the differential feature map to generate a series of weight values ​​between 0 and 1, and the output feature map is obtained through equation (6). .

[0030] (6)

[0031] In the formula, This represents the Sigmoid activation function. This indicates the addition of elements. This indicates element-wise multiplication.

[0032] As a preferred embodiment of the method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles described in this invention, the multi-receptive field feature learning consists of two symmetrical branches.

[0033] One of the branches uses and Depthwise separable convolutions are used for feature extraction in the horizontal and vertical directions, respectively.

[0034] Another branch uses a different convolution kernel for and Depth-separable convolutions are used to increase the kernel size and expand the receptive field.

[0035] As a preferred embodiment of the off-road obstacle detection method for unmanned emergency rescue vehicles described in this invention, the information interleaving consists of a Minish activation function, upsampling and downsampling operations, and a 3×3 convolution. The information interleaving is performed in the following steps:

[0036] Two feature maps are scaled to the same size using bilinear interpolation upsampling and average pooling downsampling operations.

[0037] Next, element-wise multiplication is used to fuse the two different feature maps, introducing more non-linear factors. Then, a 3×3 convolution is applied to extract the target's feature information without significantly increasing the number of parameters. Let the two input features be... , The output features are The specific expression is as follows:

[0038] (7)

[0039] (8)

[0040] In the formula, This represents the Minish activation function; This indicates a bilinear interpolation operation.

[0041] As a preferred embodiment of the off-road obstacle detection method for unmanned emergency rescue vehicles described in this invention, the multi-scale coarse-grained feature extraction branch is composed of a directional attention mechanism and information interweaving. The directional attention mechanism consists of a vertical feature extraction branch, a horizontal feature extraction branch, and a channel feature extraction branch. The multi-scale coarse-grained feature extraction branch performs the following steps:

[0042] The vertical feature extraction branch uses horizontal max pooling, average pooling, and... The convolution operation is used to deeply mine coarse-grained features in the vertical direction;

[0043] The horizontal feature extraction branch utilizes the vertical max pooling, average pooling, and... Convolutions capture contextual information in the horizontal direction;

[0044] The vertical and horizontal feature extraction branches use the Softmax activation function to normalize the extracted features, generating a series of weights to dynamically adjust the input features, achieving adaptive feature adjustment. Let the input feature map be... The outputs of the horizontal feature extraction branch and the vertical feature extraction branch are... , We obtain this through equations (9), (10), and (11);

[0045] (9)

[0046] (10)

[0047] (11)

[0048] in, This represents average pooling in the x-direction. This represents max pooling in the x-direction. This represents average pooling in the y-direction. This represents max pooling in the y-direction. This indicates maximum variance pooling. Indicates the transpose operation;

[0049] Meanwhile, the channel feature extraction branch introduces maximum variance pooling, which selects the region with the maximum variance in the feature map to highlight the features with the greatest variation. Furthermore, it uses a multilayer perceptron to perform deep modeling in the channel direction of the feature map, promoting deep fusion of information between channels. Thus, the output of the channel feature extraction branch is generated through equation (12). ;

[0050] (12)

[0051] In the formula, This represents a multilayer perceptron. Represents the ReLU activation function;

[0052] Finally, the features extracted from the three branches in different directions are fused element-wise. To prevent network degradation, residual connections are used to fuse the input feature maps. The input is used to generate the output feature map. This enhances the model's sensitivity to spatial features.

[0053] As a preferred embodiment of the method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles described in this invention, the multi-scale feature fusion strategy consists of average pooling, deconvolution, and channel stitching.

[0054] Compared with existing technologies, the beneficial effects of this invention are as follows: The core of this invention lies in its constructed CFFI part, which integrates a multi-scale fine-grained feature purification branch, a multi-scale coarse-grained feature extraction branch, and a multi-scale feature fusion strategy. Specifically, the multi-scale fine-grained feature purification branch aims to reduce environmental noise interference and deeply mine the fine-grained features of positive and negative obstacles; while the multi-scale coarse-grained feature extraction branch is mainly used to extract coarse-grained features of positive and negative obstacles at different scales; the multi-scale feature fusion strategy further fuses feature map information before inputting the feature map into the detection head, so that each feature map contains more positive and negative obstacle target feature information. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0056] Figure 1 This is an overall framework diagram of the method for identifying positive and negative obstacles in an unmanned emergency rescue vehicle under off-road conditions according to the present invention;

[0057] Figure 2 A flowchart of the multi-scale fine-grained feature purification branch provided by the present invention;

[0058] Figure 3 A flowchart illustrating the information interweaving provided for this invention;

[0059] Figure 4 A flowchart of the directional attention mechanism provided by the present invention; Detailed Implementation

[0060] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0061] A method for recognizing positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles, such as Figure 1 As shown, firstly, the Backbone is used to extract features from the input off-road environment containing positive and negative obstacles, thereby obtaining different feature layers of information for the positive and negative obstacles. The features at layers 2, 4, 6, and 9 are denoted as follows: , , , Secondly, these four feature maps are fed into the constructed Coarse-Fine Granularity Feature Interaction (CFFI) section to mitigate the information loss caused by excessively long fusion paths. CFFI includes a multi-scale fine-grained feature refinement branch, a multi-scale coarse-grained feature extraction branch, and a multi-scale feature fusion strategy (MFFS). The multi-scale fine-grained feature refinement branch is mainly used to reduce noise interference and extract multi-scale local information, such as edge details; the multi-scale coarse-grained feature extraction branch is mainly used to extract multi-scale global information, such as shape and structure; MFFS ensures that the feature maps fed into the Head section contain more target feature information. Finally, the feature maps processed by MFFS are fed into the Head section to obtain the positive and negative obstacle detection results for the off-road surface.

[0062] (1) Multi-scale fine-grained feature purification branch

[0063] The multi-scale fine-grained feature purification branch mainly consists of the constructed multi-field noise filtering feature learning (MNFFL) and information interweaving (IIW).

[0064] like Figure 2 As shown, MNFFL consists of a noise filter and multi-receptive field feature learning (MRL). Its overall process is as follows, assuming the input feature map is... The output feature map is First, use a convolution kernel size of... convolutional layers The number of channels was adjusted to increase the diversity and complexity of features; to ensure the computational cost of the algorithm, channel separation was then performed on the transformed feature map, and the feature map was obtained through equation (1). , , , And using four different sized convolution kernels respectively. , , , Multi-receptive field feature extraction is performed on four feature maps to generate feature maps. , , , Then, the Differential Attention Mechanism (DA) is applied to these feature maps to obtain the feature maps. , , , The specific expression is shown in formula (2).

[0065] (1)

[0066] (2)

[0067] In the formula, Indicates channel segmentation; This represents the differential attention mechanism; Indicates the kernel size as The convolution operation.

[0068] Secondly, for , , , Feature concatenation is performed along the channel dimension. Channel shuffling disrupts the order of feature channels, increasing the non-linearity of model learning, avoiding information redundancy, and then the concatenation is applied again. Convolution not only restores the number of channels in the feature map, but also achieves further fusion and refinement of features. The feature map is obtained through equation (3). .

[0069] (3)

[0070] In the formula, Indicates channel splicing; Indicates mixed washing of channels;

[0071] Finally, the processed feature map is added to the original feature map, which not only helps alleviate the gradient vanishing problem in deep networks, but also promotes cross-layer information interaction. The added feature map is then fed into the MRL to further extract multi-receptive field features. The output feature map is obtained through expression (4). Its specific expression is as follows:

[0072] (4)

[0073] in, This indicates multi-receptor field feature learning.

[0074] like Figure 2 As shown, DA reduces noise interference through bicubic interpolation upsampling, average pooling downsampling, and differencing operations, thereby achieving feature enhancement. Downsampling extracts background information from the feature map, while upsampling focuses on fine-grained information. The differencing operation subtracts the original image from the upsampled / downsampled feature map. Assume the input feature map is... The output feature map is Then, the difference feature map is generated through equation (5). .

[0075] (5)

[0076] In the formula, This indicates a bicubic interpolation upsampling operation; This indicates an average pooling downsampling operation; This indicates element-wise subtraction.

[0077] After differential processing, the feature map can better capture fine-grained features and reduce the interference of background noise. To further enhance the expressive power of fine-grained information, the Sigmoid activation function is used on the differential feature map to generate a series of weight values ​​between 0 and 1. These weight values ​​are then dynamically applied to the original input feature map to achieve adaptive weighting of the feature map, enabling the model to flexibly adjust the information of interest according to the importance of the features. Based on the above, the output feature map is obtained through equation (6). .

[0078] (6)

[0079] In the formula, This represents the Sigmoid activation function; This indicates element addition; This indicates element-wise multiplication.

[0080] MRL consists of two symmetrical branches. Through this symmetrical and complementary branch design, it further achieves the capture and fusion of feature information, facilitating the extraction of coarse-grained information in subsequent processes. One of the branches employs... and Depthwise separable convolutions are used for feature extraction in the horizontal and vertical directions respectively. This effectively reduces the number of parameters while enhancing the model's sensitivity to spatial dimensional features, helping to capture fine-grained features of small targets (targets with a resolution less than 32 pixels × 32 pixels). Furthermore, embedded batch normalization (BN) addresses the numerical instability problem in deep neural networks by standardizing the feature layers, making the distribution of features within the same batch more similar, thus facilitating network training. GELU is chosen as the activation function. Compared to traditional activation functions like ReLU, GELU has a smooth transition in the negative region, helping to alleviate the gradient vanishing problem and introducing more non-linearity, enhancing the model's expressive power. The feature maps fed into the MRL are denoted as... Then the output of this branch can be obtained through equation (7). .

[0081] (7)

[0082] In the formula, Indicates batch normalization; This represents the GELU activation function; Indicates that the convolution kernel is Depth-separable convolutions.

[0083] Similarly, the other branch uses a different convolution kernel. and The depthwise separable convolution, by increasing the kernel size, expands the receptive field, enabling the model to capture contextual information and facilitating the extraction of fine-grained features from large targets (targets with a resolution greater than 32 pixels × 32 pixels). This dual-branch approach is complementary: one branch focuses on fine-grained features from small targets, while the other helps capture fine-grained features from large targets. By adding the features, the outputs of the two branches are effectively fused, improving the model's ability to extract features from objects of different sizes. The output of this branch... It is obtained through equation (8).

[0084] (8)

[0085] To better utilize the features extracted by Backbone and enhance their representational power, IIW was proposed, such as... Figure 3As shown, the captured features are fused with the features extracted by the backbone using an attention mechanism. IIW mainly consists of a Minimized Activation Function, upsampling and downsampling operations, and a 3×3 convolution. The introduction of the Minimized Activation Function enhances the network's nonlinear fitting ability through its unique nonlinear mapping capability. Bilinear interpolation upsampling and average pooling downsampling operations are used to scale the two feature maps to the same size; then, element-wise multiplication is used to fuse these two different feature maps, introducing more nonlinear factors and further improving the model's expressive power. Furthermore, the 3×3 convolution effectively mines the target's feature information without significantly increasing the number of parameters. Assume the two input features are... , The output features are The specific expression is as follows:

[0086] (9)

[0087] (10)

[0088] In the formula, This represents the Minish activation function; This indicates a bilinear interpolation operation.

[0089] The fine-grained features processed by MNFFL are separated by channels to extract the feature maps compared with those extracted from the Backbone. , Feature maps with the same number of channels and Subsequently, using IIW to and respectively with , Interact with the system to generate input for the coarse-grained feature extraction branch. , This process not only refines the fine-grained feature information, but also provides diverse information for subsequent coarse-grained feature extraction, enhancing the model's expressive power and robustness.

[0090] (2) Multi-scale coarse-grained feature extraction branch

[0091] The multi-scale coarse-grained feature extraction branch mainly consists of the designed Directional Attention Mechanism (DRA) and In-place Intention (IIW). The input of the coarse-grained feature extraction branch is the output of the fine-grained feature extraction branch. , Extracted from the backbone branch .

[0092] To enhance the model's localization capabilities and facilitate the capture of coarse-grained feature information while effectively eliminating redundant information, traditional coordinate attention mechanisms have limitations in exploring feature correlations in the horizontal and vertical directions. Therefore, a directive attention (DRA) mechanism is constructed. By mining feature maps in the horizontal and vertical dimensions, it enhances the model's understanding of the spatial features of target objects, such as spatial layout and relative positional relationships.

[0093] The flowchart of DRA is as follows: Figure 4 As shown, DRA consists of three main branches: the Vertical Feature Extraction Branch (VFB), the Horizontal Feature Extraction Branch (HFB), and the Channel Feature Extraction Branch (CFB). VFB utilizes max pooling, average pooling, and... Convolutional operations are used to deeply mine coarse-grained features in the vertical direction. Similarly, HFB utilizes max pooling, average pooling, and... The convolutions capture horizontal contextual information. Simultaneously, both branches utilize a reshape operation to uncover potential correlations between different features, enriching the coarse-grained information of the feature map. To further enhance feature representation capabilities, the VFB and HFB branches employ the Softmax activation function to normalize the extracted features, generating a series of weights to dynamically adjust the input features, achieving adaptive feature adjustment. Assume the input feature map is... Then the outputs of VFB and HFB , It can be generated by equations (11), (12) and (13).

[0094] (11)

[0095] (12)

[0096] (13)

[0097] in, This represents average pooling in the x-direction. This represents max pooling in the x-direction. This represents average pooling in the y-direction. This represents max pooling in the y-direction. This indicates maximum variance pooling. This indicates the transpose operation.

[0098] Meanwhile, CFB introduces Max Variance Pooling, which selects the region with the largest variance in the feature map to highlight the most variable features and effectively extract coarse-grained information in the channel direction. Furthermore, a Multilayer Perceptron (MLP) is used to perform deep modeling in the channel direction of the feature map, promoting deep fusion of information between channels and improving the richness of feature information. The output of CFB is then generated using Equation (14). .

[0099] (14)

[0100] In the formula, This represents a multilayer perceptron. This represents the ReLU activation function.

[0101] Furthermore, the features extracted from the three branches in different directions are fused using element-wise addition. Simultaneously, to prevent network degradation, the input feature maps are processed using residual connections. The input is used to generate the output feature map. This enhances the model's sensitivity to spatial features.

[0102] (3) Multi-scale feature fusion strategy

[0103] In the traditional YOLOv8 architecture, the feature maps after Neck fusion are directly fed into the detection head. However, these feature maps often lack a sufficient balance between local details and global information, failing to fully encompass both fine-grained and coarse-grained feature information of the target. To improve the detection performance for targets of different sizes, MFFS is proposed to ensure that the feature maps contain more fine-grained and coarse-grained information about the target.

[0104] The structure diagram of MFFS is as follows Figure 1As shown in the last dashed box in the GFFI diagram, it mainly consists of average pooling, deconvolution, and channel concatenation. Average pooling, as a downsampling method, effectively captures and enhances the overall features and coarse-grained information of the feature map by reducing its spatial size. Next, deconvolution is used for upsampling to reduce the spatial resolution loss caused by average pooling. Deconvolution improves the spatial resolution of the feature map by learning the relationships between neighboring pixels, enhancing the expressive power of fine-grained features while preserving the original information. By alternately using downsampling and upsampling, a progressively guided mechanism is formed. In this process, high-resolution feature maps (rich in fine-grained information) and low-resolution feature maps (rich in coarse-grained information) are continuously fused, with each fusion based on further optimization of the previous fusion result. Through this progressive refinement, feature maps containing rich coarse-grained and fine-grained information are generated, thereby further improving detection performance. The overall process is shown below:

[0105] (15)

[0106] (16)

[0107] in, Indicates that the convolution kernel is The deconvolution.

[0108] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for detecting positive and negative obstacles on off-road surfaces by an unmanned emergency rescue vehicle, characterized in that, The steps are as follows: S1. First, the YOLOv8 Backbone is used to extract features from the input off-road environment containing positive and negative obstacles, thereby obtaining different feature layers of positive and negative obstacles. The features at layers 2, 4, 6, and 9 are denoted as follows: , , , ; S2. Next, the four feature maps are sequentially fed into the constructed multi-scale fine-grained feature purification branch, multi-scale coarse-grained feature extraction branch, and multi-scale feature fusion strategy. The multi-scale fine-grained feature purification branch is used to reduce noise interference and extract multi-scale local information, the multi-scale coarse-grained feature extraction branch is used to extract multi-scale global information, and the multi-scale feature fusion strategy makes the feature map fed into the Head part contain more target feature information. S3. Finally, the feature map processed by the multi-scale feature fusion strategy is fed into the Head part to obtain the positive and negative obstacle detection results on the off-road surface.

2. The method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles according to claim 1, characterized in that, The multi-scale fine-grained feature purification branch consists of multi-receptive-field noise filtering feature learning and information interleaving.

3. The method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles according to claim 2, characterized in that, The multi-receptive-field noise filtering feature learning consists of a noise filter and multi-receptive-field feature learning. Multi-receptive-field noise filtering feature learning is performed using the following steps: Let the input feature map be The output feature map is First, use a convolution kernel size of convolutional layers The number of channels was adjusted, and then the channels of the transformed feature map were separated. The feature map was obtained through equation (1). , , , And using four different sized convolution kernels respectively. , , , Multi-receptive field feature extraction is performed on four feature maps to generate feature maps. , , , ; Next, the differential attention mechanism is applied to these feature maps using formula (2) to obtain the feature maps. , , , ; （1）； （2）； In the formula, Indicates channel segmentation; This represents the differential attention mechanism; Indicates the kernel size as Convolution operations; Secondly, for , , , Feature concatenation is performed along the channel dimension, and the order of feature channels is shuffled through a channel shuffling operation before being applied again. Convolution, through equation (3), yields the feature map. ; （3）； In the formula, Indicates channel splicing. Indicates mixed washing of channels; Finally, the processed feature map is added to the original feature map, which not only helps to alleviate the gradient vanishing problem in deep networks, but also promotes cross-layer information interaction; the added feature map is then fed into MRL to further extract multi-receptive field features. The output feature map is obtained through expression (4). ; （4）； in, This indicates multi-receptor field feature learning.

4. The method for detecting positive and negative obstacles on off-road surfaces for an unmanned emergency rescue vehicle according to claim 3, characterized in that, The differential attention mechanism performs the following steps: Let the input feature map be The output feature map is Then, the difference feature map is generated through equation (5). ; （5）； In the formula, This indicates a bicubic interpolation upsampling operation; This indicates an average pooling downsampling operation; This indicates element-wise subtraction. After differential processing, the Sigmoid activation function is applied to the differential feature map to generate a series of weight values ​​between 0 and 1, and the output feature map is obtained through equation (6). ; （6）； In the formula, This represents the Sigmoid activation function. This indicates the addition of elements. This indicates element-wise multiplication.

5. The method for detecting positive and negative obstacles on off-road surfaces for an unmanned emergency rescue vehicle according to claim 3, characterized in that, The multi-receptive-field feature learning consists of two symmetrical branches; One of the branches uses and Depthwise separable convolutions are used for feature extraction in the horizontal and vertical directions, respectively. Another branch uses a different convolution kernel for and Depth-separable convolutions are used to increase the kernel size and expand the receptive field.

6. The method for detecting positive and negative obstacles on off-road surfaces for unmanned emergency rescue vehicles according to claim 2, characterized in that, The information interleaving consists of a Minish activation function, upsampling and downsampling operations, and a 3×3 convolution. The information interleaving is performed in the following steps: Two feature maps are scaled to the same size using bilinear interpolation upsampling and average pooling downsampling operations. Next, element-wise multiplication is used to fuse the two different feature maps, introducing more non-linear factors. Then, a 3×3 convolution is applied to extract the target's feature information without significantly increasing the number of parameters. Let the two input features be... , The output features are The specific expression is as follows: （7）； （8）； In the formula, This represents the Minish activation function; This indicates a bilinear interpolation operation.

7. The method for detecting positive and negative obstacles on off-road surfaces for an unmanned emergency rescue vehicle according to claim 1, characterized in that, The multi-scale coarse-grained feature extraction branch is composed of a directional attention mechanism and information interweaving. The directional attention mechanism consists of a vertical feature extraction branch, a horizontal feature extraction branch, and a channel feature extraction branch. The multi-scale coarse-grained feature extraction branch performs the following steps: The vertical feature extraction branch uses horizontal max pooling, average pooling, and... The convolution operation is used to deeply mine coarse-grained features in the vertical direction; The horizontal feature extraction branch utilizes the vertical max pooling, average pooling, and... Convolutions capture contextual information in the horizontal direction; The vertical and horizontal feature extraction branches use the Softmax activation function to normalize the extracted features, generating a series of weights to dynamically adjust the input features, achieving adaptive feature adjustment. Let the input feature map be... The outputs of the horizontal feature extraction branch and the vertical feature extraction branch are... , We obtain this through equations (9), (10), and (11); （9）； (10)； (11)； in, This represents average pooling in the x-direction. This represents max pooling in the x-direction. This represents average pooling in the y-direction. This represents max pooling in the y-direction. This indicates maximum variance pooling. Indicates the transpose operation; Meanwhile, the channel feature extraction branch introduces maximum variance pooling, which selects the region with the maximum variance in the feature map to highlight the features with the greatest variation. Furthermore, it uses a multilayer perceptron to perform deep modeling in the channel direction of the feature map, promoting deep fusion of information between channels. Thus, the output of the channel feature extraction branch is generated through equation (12). ; （12）； In the formula, This represents a multilayer perceptron. Represents the ReLU activation function; Finally, the features extracted from the three branches in different directions are fused element-wise. To prevent network degradation, residual connections are used to fuse the input feature maps. The input is used to generate the output feature map. This enhances the model's sensitivity to spatial features.

8. The method for detecting positive and negative obstacles on off-road surfaces for an unmanned emergency rescue vehicle according to claim 1, characterized in that, The multi-scale feature fusion strategy consists of average pooling, deconvolution, and channel concatenation.

Images