Battlefield medical unmanned aerial vehicle obstacle avoidance recognition method and device based on artificial intelligence

Abstract

The invention relates to a battlefield medical unmanned aerial vehicle obstacle avoidance recognition method and device based on artificial intelligence, and belongs to the technical field of artificial intelligence and data processing. The method comprises the following steps: acquiring battlefield image data and marking the battlefield image data to obtain a battlefield image data set; performing adaptive illumination normalization and dynamic defogging enhancement on the images in the data set to obtain an enhanced battlefield image data set; a convolutional neural network model is constructed, the model comprises three different branches and a feature fusion module, the branch 2 adopts a dynamic kernel generation mechanism in the convolution process, and the feature fusion module introduces an attention gating mechanism; training a convolutional neural network model through the enhanced battlefield image; optimizing a training process by adopting a loss function to obtain a trained model; and inputting a battlefield image to be detected into the trained model, and identifying and classifying different obstacles in the image. The accuracy and robustness of battlefield medical unmanned aerial vehicle obstacle avoidance recognition and classification can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and data processing technology, and specifically relates to an artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method and equipment. Background Art

[0002] In modern battlefield environments, medical drones are widely used for emergency rescue and casualty transport missions. Their efficiency and timeliness are crucial to the survival of casualties on the battlefield. However, in complex and ever-changing battlefield environments, drones face numerous challenges, one of the most critical being how to accurately identify and avoid obstacles in real time. Obstacles come in many forms, such as buildings, casualties, explosives, smoke, and flames. These obstacles often appear in extreme lighting conditions, smoke obscuration, or in dynamic environments. This requires drones to possess robust image processing and obstacle avoidance capabilities when performing their missions.

[0003] Existing methods incorporate deep learning algorithms to analyze drone-captured images of construction pavement to predict the density of rockfill and guide the roller's path, thereby improving construction efficiency and compaction quality. Historical radar data is used to preprocess and label target objects, and real-time radar data is fed into the drone's target recognition model based on priority sorting to improve recognition efficiency and accuracy. Furthermore, the drone's onboard environmental data acquisition module acquires wind speed and light intensity, calculates the number of images to be taken, and then captures targets such as transformers. Finally, image recognition terminals are used for recognition, adapting to different spatial environments and reducing recognition error rates. These methods have collectively promoted the development of drone obstacle avoidance and recognition technology, improving the intelligence level and application effectiveness of drones. However, the following problems still need to be further addressed: Most existing technologies use fixed convolution kernels, which are difficult to cope with obstacles with huge scale changes in battlefield images, resulting in a fixed network perception range and difficulty in adapting to the distribution density and size changes of different obstacles; traditional multi-branch feature fusion methods often directly splice the outputs of different branches, which easily leads to the introduction of redundant information and cannot automatically adjust the contribution ratio of branches according to different scenarios; traditional methods often rely on global normalization or simple dehazing algorithms when dealing with environments with drastic changes in lighting or smoke obscuration, which leads to unstable image quality and affects the effect of feature extraction; when faced with complex battlefield scenes, traditional methods are often unable to handle the interference of various types of obstacles and backgrounds, resulting in large differences in the performance of the model in different environments. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides an artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method and equipment.

[0005] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: The present invention provides an artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method, comprising the following steps: S1. Obtain battlefield image data and annotate it to obtain a battlefield image dataset; S2. Perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; S3. Construct a convolutional neural network model, comprising three different branches and a feature fusion module. Branch 1 uses a 1×1 convolution kernel; branch 2 uses a 3×3 convolution kernel with a dynamic kernel generation mechanism during the convolution process; branch 3 uses identity mapping; the feature fusion module incorporates an attention gating mechanism; the convolutional neural network model is trained using images from the enhanced battlefield image dataset; and the training process is optimized using a loss function to obtain a trained model. S4. The battlefield image to be detected is input into the trained model to identify and classify different obstacles in the image.

[0006] Furthermore, in step S1, image data from different battlefield environments are collected. The image data should cover various complex obstacles and environmental conditions, including buildings, wounded people, explosion fragments, smoke, and fire; the labeled categories include building debris, wounded people, explosives, and drones.

[0007] Furthermore, step S2 specifically includes: An adaptive dehazing algorithm based on the atmospheric scattering model is used to process images in the battlefield image dataset. Combined with dynamic normalization of illumination compensation, an enhanced battlefield image dataset is obtained by dynamically estimating atmospheric light and transmittance. The formula is as follows: , in, Indicates the image in the enhanced battlefield image dataset at coordinates The pixel value at ; Indicates the coordinates of the image in the battlefield image dataset The pixel value at ; Indicates the channel adaptive atmospheric light estimation Channel Position pixel value, Represents the coordinates in the transmittance map The transmittance value at express The local illumination compensation coefficient of the position, Represents a constant.

[0008] Furthermore, in the calculation In the process of dynamic estimation, the smoke concentration of each channel is dynamically estimated by combining the dark channel prior and the exponential weighted average of the local window.

[0009] Furthermore, the 3×3 convolution process of branch 2 in step S3 adopts a dynamic kernel generation mechanism to dynamically adjust the convolution weights according to the input features, so that the model automatically increases the receptive field in dense obstacle areas and focuses on local details in small target areas.

[0010] Furthermore, the feature fusion module in step S3 adopts a feature fusion method based on the attention gating mechanism, dynamically weights the contribution of each branch, and calculates the channel-level correlation so that the model relies on the original features of branch 3 when the scene is occluded by smoke, while the clear scene focuses on the deep features of branches 2 and 1.

[0011] Furthermore, the loss function in step S3 uses contrast loss to enhance intra-class compactness. By adopting the transmittance-aware contrast loss term and dynamically adjusting the contrast loss weight in combination with smoke density and illumination changes, the model can achieve robust feature learning under smoke occlusion and illumination changes, forcing obstacles of the same type to cluster in the feature space and separate heterogeneous objects.

[0012] Furthermore, during the calculation of the loss function in step S3, in the calculation of the transmittance perception contrast loss term, a dynamic weighting mechanism of the transmittance and illumination perception contrast loss is used to achieve the weakened constraint of the smoke area and the enhanced separation of the high illumination area.

[0013] Furthermore, in step S3, the convolutional neural network model uses a transmittance-channel joint gated classifier function to dynamically reconstruct the classification weights by fusing features with the cross-modal interaction of the transmittance map and the illumination coefficient, and finally outputs the category probability, specifically including: Transmittance-aware channel enhancement: A learnable smoke-sensitive channel selection matrix is ​​constructed, and high-frequency compensation weights are calculated through cross-attention between the transmittance map and feature channels. Illumination-robust channel suppression: Designing an illumination-channel covariance matrix to suppress noisy channel responses under high illumination variations, and calculating the correlation between illumination coefficients and feature global statistics through outer products to achieve channel-level dynamic suppression. The enhanced and suppressed features are interacted cross-modally to generate a dynamic classification weight matrix. The transmittance and illumination perception features are fused through bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression to finally output the category probability.

[0014] The present invention also provides an artificial intelligence-based battlefield medical drone obstacle avoidance and recognition device, which executes the above-mentioned artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method, including: A battlefield image acquisition unit is used to acquire battlefield image data and annotate it to obtain a battlefield image data set; Graphics enhancement unit, used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; A model training unit is used to build a convolutional neural network model and train the convolutional neural network model using images from the enhanced battlefield image dataset. The training process is optimized using a loss function to obtain a trained model. The result generation unit is used to obtain the image to be detected, and input the battlefield image to be detected into the trained model to identify and classify different obstacles in the image.

[0015] The advantages of the present invention are: The present invention adopts an adaptive dehazing algorithm based on the atmospheric scattering model, combined with a dynamic normalization method of illumination compensation, which can dynamically estimate atmospheric light and transmittance, thereby solving the influence of smoke interference and illumination changes on image features, significantly enhancing image quality, and making subsequent feature extraction more accurate; by designing a multi-branch convolutional neural network model, combining convolution kernels of different sizes and identity mapping, the spatial features and local details of the image are effectively extracted. The multi-branch structure can adaptively handle complex obstacles and enhance the robustness of the network in changing scenarios. By using identity mapping branches to retain the original information, the interference of noise on key features is avoided; traditional convolutional neural networks usually use fixed convolution kernels. The present invention proposes a dynamic convolution kernel generation mechanism, which dynamically adjusts the convolution kernel weights according to changes in input features, so that the model adapts to multi-scale obstacles and improves performance in complex battlefield environments, such as automatically enhancing features in high-density or large-size obstacle areas. Extraction capability; The present invention proposes a feature fusion method based on the attention gating mechanism, which can automatically adjust the feature contribution of different convolution branches according to the complexity of different scenes, avoid the feature redundancy and noise interference that may be caused by the traditional feature splicing method, and optimize the fusion effect of image features; the transmittance-aware contrast loss is adopted to enhance the intra-class compactness and inter-class separation, especially in the case of smoke occlusion and illumination changes, by dynamically adjusting the weight of the loss function, the obstacle recognition accuracy is optimized; in addition, the illumination compensation mechanism improves the image quality under high illumination conditions and enhances the robustness of the system by dynamically calculating the global and local illumination means; the proposed transmittance-channel joint gated classifier function fuses the transmittance map and illumination information with the image features through cross-modal interaction, and dynamically adjusts the classification weights, which can enhance information retention in the smoke area and enhance the inter-class separation in the area with strong illumination, thereby improving the accuracy and robustness of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0017] Figure 1 This is a diagram of the architecture of the method of the present invention; Figure 2 This is a diagram showing the spatial adaptability analysis of the dynamic convolution kernel of the present invention; Figure 3 This is the attention gate feature fusion analysis diagram of the present invention; Figure 4 This is a performance analysis diagram for verifying the transmittance perception loss of the present invention; Figure 5 This is a performance verification diagram of the illumination compensation mechanism of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] Example 1 In this embodiment, Figure 1 As shown, a battlefield medical drone obstacle avoidance and recognition method based on artificial intelligence is provided, and the specific steps include: S1. Obtain battlefield image data and annotate it to obtain a battlefield image dataset; Specifically, battlefield medical drone obstacle avoidance and recognition tasks require the collection of image data from diverse battlefield environments. This image data should cover a wide range of scenarios, including complex obstacles and environmental conditions such as buildings, casualties, explosion debris, smoke, and flames. Because battlefield images are typically highly diverse, including varying lighting conditions, obstructions, and dynamic objects, data collection must ensure comprehensiveness and diversity, encompassing a wide range of possible battlefield scenarios. During the collection process, it is crucial to ensure that each image is representative enough to capture the diverse battlefield obstacles and their interactions with the surrounding environment.

[0020] To annotate battlefield image data, every obstacle and key feature in the image must be accurately labeled, ensuring that each image sample is associated with a corresponding label. Annotations typically include the obstacle's category, location, and relative positional relationships. Annotated categories include building debris, casualties, explosives, and drones. This annotation work is performed manually by professionals to ensure accuracy and consistency.

[0021] S2. Perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; Specifically, battlefield image data has the characteristics of drastic lighting changes, smoke occlusion and dynamic motion blur. At the same time, battlefield images are usually affected by smoke occlusion and lighting changes, resulting in image feature degradation; Traditional normalization only linearly scales pixel values ​​and cannot eliminate the interference of smoke and motion blur on texture features; An adaptive dehazing algorithm based on an atmospheric scattering model is used to process images in the battlefield image dataset. Combined with dynamic normalization of illumination compensation, an enhanced battlefield image dataset is obtained by dynamically estimating atmospheric light and transmittance, solving the problem of feature degradation under smoke interference. The formula is as follows: , in, Indicates the image in the enhanced battlefield image dataset at coordinates The pixel value at ; Indicates the coordinates of the image in the battlefield image dataset The pixel value at ; Indicates the channel adaptive atmospheric light estimation Channel Position pixel value, express The local illumination compensation coefficient of the position, Indicates a constant used to avoid division by zero errors and is set to 0.0001; Specifically, in the calculation In the process of combining dark channel prior, the smoke concentration of each channel is dynamically estimated through the exponential weighted average of the local window. The calculation formula is as follows: , It should be noted that battlefield smoke distribution is uneven, such as the uneven distribution of local smoke from explosions. Suppress the interference of non-smog areas and enhance the estimation accuracy of smoke areas; Represents a local window, i.e., a region defined in an image. Its size is set according to specific application requirements, and window sizes such as 3×3 and 5x5 can be selected. This window is used to calculate the statistical characteristics of local pixel values ​​to help derive the local brightness and transmittance of the image; Represents the pixel coordinates within the local window, represents an exponential function with a natural constant as the base, represents the L2 norm; represents the dark channel prior, and the formula is as follows: , in, Indicates the minimum operation of all pixel coordinates in the local window, Represents the image in the battlefield image dataset Channel Position pixel value; Represents the dark channel smoothing coefficient, set to ; Represents the coordinates in the transmittance map The transmittance value at is expressed as follows: , in, Indicates the smoke density adjustment factor. The smoke density varies greatly in the battlefield. By adjusting the smoke density adjustment factor, the balance between defogging and detail retention is achieved. Set it to ; Indicates the minimum value among the red, green, and blue channels. It is used to represent the minimum color channel information at that point to estimate the dark channel of the image and adapt to the non-uniform attenuation of RGB channels by smoke in battlefield images. For example, the red channel is easily interfered by firelight. Represents the image in the battlefield image dataset Channel Position pixel value, The channel can be red, green, or blue. By calculating the ratio of the global mean to the local mean, the drastic changes in battlefield illumination can be dynamically compensated. The formula is as follows: , in, represents the global pixel mean, represents the sliding window mean; Represents a constant used to avoid division by zero errors and is set to 0.0001.

[0022] S3. Construct a convolutional neural network model, comprising three different branches and a feature fusion module. Branch 1 uses a 1×1 convolution kernel; branch 2 uses a 3×3 convolution kernel with a dynamic kernel generation mechanism during the convolution process; branch 3 uses identity mapping; the feature fusion module incorporates an attention gating mechanism; the convolutional neural network model is trained using images from the enhanced battlefield image dataset; and the training process is optimized using a loss function to obtain a trained model. Branch 1 uses a 1×1 convolution kernel to change the number of channels in the feature map, reduce the dimension or increase the depth of the network, and retain spatial information. Battlefield images may contain a large amount of redundant information. When the amount of data is large, 1×1 convolution can effectively reduce the dimension and avoid information overload. In one embodiment, if the input battlefield image data size is 3×3, then a 3×3 feature map is obtained through a 1×1 convolution kernel. The convolution operation of multiple 1×1 convolution kernels helps to refine the feature map by changing the number of channels. Similarly, if the input battlefield image data size is 1024x1024, then a 1×1 convolution kernel is used to obtain a 1024x1024 feature map.

[0023] Traditional multi-branch networks use fixed convolution kernels, which make it difficult to adapt to the multi-scale characteristics of battlefield obstacles, such as the size differences of explosion fragments, wounded limbs, and building debris. Branch 2 uses dynamic convolution kernel weights to determine the convolution kernel weight parameters of branch 2 and uses padding operations to keep the spatial dimension of the feature map unchanged and extract spatial features. In battlefield images, complex obstacles such as buildings and wounded people have large scale variations. Dynamic convolution kernels can adaptively process obstacles of different sizes and enhance adaptability to different types of targets. Specifically, a dynamic kernel generation mechanism is used in the 3×3 convolution of branch 2. The convolution weights are dynamically adjusted according to the input features, so that the model automatically increases the receptive field in dense obstacle areas and focuses on local details in small target areas. The formula is as follows: , , in, Indicates branch 2 The initial convolution kernel weights of the layer; Represents the dynamic convolution kernel weight of the k-th layer of branch 2. According to the changes in input features, the convolution kernel weight will be dynamically adjusted to adapt to changes in different scenarios; Indicates branch 2 The convolution kernel weight increment of each layer is dynamically adjusted based on input features during training. The dynamic calculation of convolution kernel weights can adaptively adjust the convolution kernel according to the input features, solving the limitations of traditional fixed convolution kernels when dealing with multi-scale obstacles. For example, when the input is a smoke-occluded area, the global average pooling operation outputs a low activation value, and the fully connected network reduces the convolution kernel weight, reducing noise propagation. Represents the global average pooling operation, which reduces the dimension of the feature map by calculating the average value of the entire feature map, thereby obtaining features with more global information; is the input feature map of the kth layer of branch 2, which represents the feature representation before the convolution operation of this layer; It is a fully connected network with the LeakyReLU activation function. By processing the features after the global average pooling operation, the network can automatically adjust the size of the receptive field according to the different battlefield environments, thereby focusing on details in dense areas and maintaining broad perception in sparse areas. Branch 3 is identity mapping. This involves passing the input feature map directly through the network, leaving it unchanged. This preserves the original information and mitigates the vanishing gradient problem. In battlefield imagery, certain features can be very weak and easily lost due to factors like noise and occlusion. Identity mapping provides a direct transfer of the original information, ensuring that the network does not lose critical information when processing complex images.

[0024] In battlefield images, smoke and occlusion may cause information loss in some areas. Conventional methods usually directly splice the outputs of the three branches, which easily leads to feature redundancy and noise sensitivity. Specifically, the feature fusion module adopts a feature fusion method based on the attention gating mechanism, dynamically weighting the contribution of each branch. Through channel-level correlation calculation, the model relies on the original features of branch 3 in smoke occlusion, while focusing on the deep features of branches 2 and 1 in clear scenes. The formula is expressed as follows: , , in, Represents the Sigmoid activation function; It represents a compression-excitation module consisting of two 1×1 convolutional layers, which maps the channel-concatenated features into a three-dimensional attention weight space; For the Attention weights of branches: Through the attention gating mechanism, the model dynamically adjusts the weights of each branch to better integrate features of different scales and levels; It is the feature map output by branch 1; It is the feature map output by branch 2; It is the feature map output by branch 3; Express 、 and Perform channel splicing operations; It is a fused multi-scale feature map. Through dynamic weighting, it ensures that it relies more on the original information when the occlusion is severe, and relies more on the deep features when the image is clear. is the attention weight of branch 1; is the attention weight of branch 2; is the attention weight of branch 3; Represents channel-by-channel multiplication.

[0025] Specifically, during model training, the conventional single cross-entropy loss cannot distinguish between similar obstacle categories. For example, a wounded person and a sandbag can easily be considered similar obstacles. Contrast loss is used to enhance intra-class compactness. By adopting a transmittance-aware contrast loss term and dynamically adjusting the contrast loss weight in combination with smoke density and illumination changes, the model achieves robust feature learning under smoke occlusion and illumination changes, forcing obstacles of the same type to cluster in the feature space and separating different types. The formula is as follows: , in, It is a contrast loss function that improves the recognition ability of the model in multi-obstacle environments by strengthening intra-class compactness and inter-class separation. For example, in a smoke area, the transmittance-aware contrast loss helps distinguish the characteristics of different obstacles. is the number of training samples in a single batch; For the A set of positive samples of the same type; The feature vector extracted by the multi-branch convolutional neural network for the i-th sample; is the battlefield image sample input to the multi-branch convolutional neural network; The feature vector extracted by the multi-branch convolutional neural network for the p-th sample (similar samples); is the pth battlefield image sample input to the multi-branch convolutional neural network, and is the feature vector of similar samples; The feature vector extracted by the multi-branch convolutional neural network for the nth sample (heterogeneous sample); The nth battlefield image sample input to the multi-branch convolutional neural network is the heterogeneous sample feature vector; For the A negative sample set of different classes of samples; is the balance coefficient; , which means that the feature space aggregation of similar samples is achieved through feature distance constraint; To maximize the function, negative values ​​are truncated to zero, ensuring that the contrast loss only generates gradient updates when the feature distance of similar samples is greater than the feature distance of heterogeneous samples; is the transmittance perceived contrast loss term.

[0026] Specifically, to address the problem of feature blurring caused by smoke occlusion and inter-class confusion under drastic lighting changes, a dynamic weighting mechanism of transmittance and lighting perception contrast loss is used during the calculation of the transmittance-perceived contrast loss term to achieve weakened constraints in smoke areas and enhanced separation in high-light areas. The formula for the transmittance-perceived contrast loss term is expressed as follows: , in, represents the mean value of the transmittance map; It represents the smoke area weakening contrast item, and the calculation method is expressed as ,When the smoke is dense, that is, when the transmittance is low, the distance constraint of similar sample features is reduced to avoid forcing the model to learn invalid features under smoke occlusion; For samples The mean transmittance of the corresponding area represents the weighted positive sample comparison item by transmittance, and the smoke-occluded area Reduce feature aggregation constraints, i.e., , for local area; is the coordinate in the transmittance diagram The transmittance value at is calculated as ; The image in the battlefield image dataset is Channel Position pixel value, The channel can be red, green, or blue; is the first in channel-adaptive atmospheric light estimation. Channel Position pixel value, representing the adaptive atmospheric light value of each channel; Represents the cardinality of the positive sample set; is the illumination-sensitive negative sample reinforcement term, and its calculation method is expressed as When the illumination changes dramatically, that is, when the local illumination compensation coefficient is large, the separation strength of heterogeneous samples is enhanced to prevent inter-class confusion in the highlight area, such as preventing inter-class confusion between fire and orange obstacles; For samples The mean value of the illumination compensation coefficient, , enhance the inter-class separation strength in areas with high illumination changes; for The local illumination compensation coefficient of the location; Represents the cardinality of the negative sample set; is the variance of the transmittance map; is the global mean of the transmittance map, and the calculation method is expressed as , and are the height and width of the transmittance map, respectively; Characterizing the ratio of the variance to the mean of the transmittance map, it forces the model to automatically balance the loss weights when the smoke is unevenly distributed.

[0027] Specifically, the convolutional neural network model uses a transmittance-channel joint gated classifier function to dynamically reconstruct classification weights by fusing features with cross-modal interactions of transmittance maps and illumination coefficients, and ultimately outputs category probabilities. The model includes the following steps: (1) Transmittance perception channel enhancement: A learnable smoke-sensitive channel selection matrix is ​​constructed, and the high-frequency compensation weight is calculated through cross-attention between the transmittance map and the feature channel. The formula is as follows: , in, Enhance the attention weight for the channel, the dimension is , used to select smoke-sensitive feature channels, and compensate for the attenuation of high-frequency information caused by smoke through cross-attention between the transmittance map and the feature channel. That is, the channel weight corresponding to the smoke area with low transmittance is reduced, reducing the noise channel response; It is the Softmax function, which performs normalization along the channel dimension; is the fused multi-scale feature map; for The transpose of It is the global maximum pooling operation; is the global average pooling operation; is a learnable spatial transmittance adaptation matrix; is the number of channels of the fused multi-scale feature map.

[0028] (2) Illumination robust channel suppression: Design the illumination-channel covariance matrix to suppress the noisy channel response under high illumination changes. Calculate the correlation between the illumination coefficient and the feature global statistics through the outer product to achieve channel-level dynamic suppression. The formula is as follows: , in, is the channel suppression coefficient; represents the outer product of tensors, represents the local illumination compensation coefficient, Indicates that and Perform outer product to generate covariance features to suppress channel noise caused by drastic changes in illumination, such as channel noise caused by overexposed areas under strong light.

[0029] (3) The enhanced and suppressed features are cross-modally interacted to generate a dynamic classification weight matrix. The transmittance and illumination perception features are fused through bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression. The final output category probability is expressed as follows: , , , , in, is the characteristic map after transmittance enhancement; is the feature map after light suppression; Represents channel splicing operation; The dynamically generated classification weight matrix is ​​a learnable parameter that maps the global statistics of transmittance and illumination to the classification space through a multi-layer perceptron. for The transpose of is the classification bias, which is a learnable parameter; is the final category probability distribution, express Types of obstacles, such as injured people, building debris, drones, etc. is the Softmax function.

[0030] S4. Input the battlefield image to be detected into the trained model to identify and classify different obstacles in the image.

[0031] Example 2 In this embodiment, dynamic convolution kernel spatial adaptability analysis is performed. By constructing a three-dimensional spatial surface, the influence of obstacle distribution density and size on detection accuracy is explored. Figure 2 As shown, compared with the traditional fixed convolution kernel Figure 2 (a) Dynamic convolution kernel with this technology Figure 2 (b) Adaptability in different battlefield environments. Experimental results show that when the obstacle density and size in the battlefield environment are high, the average accuracy surface of the traditional method shows a significant depression, indicating that its fixed receptive field is difficult to adapt to the complex and changeable obstacle distribution. The surface generated by this technology remains stable, and the color mapping display accuracy is always maintained at a high level. This verifies that the dynamic convolution kernel can adaptively expand the perception range according to the input features through the dynamic adjustment mechanism of weight increments, automatically enhance the feature extraction capability in dense and large-scale obstacle scenes, and overcome the limitations of traditional methods in complex battlefield environments.

[0032] Example 3 In this embodiment, the working principle of the multi-branch feature fusion mechanism is revealed by comparing the branch weight distribution in clear scenes and smoke-occluded scenes. The stacked bar chart corresponding to the attention gate feature fusion analysis is shown in Figure 2. Figure 3 As shown, Figure 3 (a) is the attention weight distribution of the clear scene, Figure 3(b) is the distribution of attention weights for smoke occlusion. In a smoke-filled environment, the weight of the third branch representing the retention of original information is significantly improved, while the dynamic convolution branch dominates in clear scenes. The dynamic adjustment stems from the real-time perception of battlefield environment characteristics by the attention gating mechanism. When the image quality degrades due to smoke, the system automatically enhances the contribution of the identity mapping branch to ensure that key obstacle information is not overwhelmed by noise. The experimental results intuitively demonstrate the intelligent feature selection capability achieved by this technology through channel-level weight allocation, solving the problem of low information utilization in traditional methods in occluded scenes.

[0033] Example 4 like Figure 4 As shown in the figure, this embodiment verifies the effectiveness of transmittance perception contrast loss by analyzing the curve of the impact of changes in smoke concentration on obstacle recall. The experimental curve shows that the recall rate of the traditional method shows a sharp and oscillating decline with increasing smoke concentration, and the confidence interval range gradually expands. However, the curve generated by this technology declines smoothly with a small fluctuation amplitude. The performance stability in high-concentration areas is significantly better than the comparison method. Because this technology uses a transmittance variance adjustment mechanism to automatically reduce invalid feature constraints in areas with uneven smoke distribution, while strengthening the inter-class discrimination of effective features. The comparison of the shaded areas in the figure proves that the dynamic correlation model between smoke concentration and loss weight established by this technology effectively alleviates the feature confusion problem of traditional methods in dense smoke environments.

[0034] Example 5 This embodiment uses a dual-axis graph to comprehensively evaluate the impact of light intensity changes on the detection system. Figure 5 (a) shows that the accuracy of the traditional method exhibits a parabolic decay as the light intensity increases, while the curve of this technology only shows slight fluctuations under extreme light conditions; Figure 5 The error distribution box plot in (b) further shows that the error dispersion of this technology is reduced by more than 50% compared with traditional methods, indicating that the dynamic correction of local brightness differences by the illumination compensation coefficient, calculated by the ratio of the global and local illumination means, effectively suppresses pixel distortion in overexposed areas. The performance comparison on the logarithmic coordinate axis in the figure shows that the illumination-transmittance joint compensation model constructed by this technology significantly improves the robustness of the system under complex lighting conditions on the battlefield.

[0035] Example 6 This embodiment provides an artificial intelligence-based battlefield medical drone obstacle avoidance and recognition device, which executes the artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method described in Example 1, including: A battlefield image acquisition unit is used to acquire battlefield image data and annotate it to obtain a battlefield image data set; Graphics enhancement unit, used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; A model training unit is used to build a convolutional neural network model and train the convolutional neural network model using images from the enhanced battlefield image dataset. The training process is optimized using a loss function to obtain a trained model. The result generation unit is used to obtain the image to be detected, and input the battlefield image to be detected into the trained model to identify and classify different obstacles in the image.

[0036] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method, characterized in that: The following steps are involved: S1. Obtain battlefield image data and annotate it to obtain a battlefield image dataset; S2. Perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; S3. Construct a convolutional neural network model, comprising three different branches and a feature fusion module. Branch 1 uses a 1×1 convolution kernel; branch 2 uses a 3×3 convolution kernel with a dynamic kernel generation mechanism during the convolution process; branch 3 uses identity mapping; the feature fusion module incorporates an attention gating mechanism; the convolutional neural network model is trained using images from the enhanced battlefield image dataset; and the training process is optimized using a loss function to obtain a trained model. S4. The battlefield image to be detected is input into the trained model to identify and classify different obstacles in the image.

2. The battlefield medical drone obstacle avoidance and recognition method based on artificial intelligence according to claim 1 is characterized in that: In step S1, image data from different battlefield environments are collected. The image data should cover various complex obstacles and environmental conditions, including buildings, wounded people, explosion fragments, smoke, and fire; the labeled categories include building debris, wounded people, explosives, and drones.

3. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 2 is characterized in that: Step S2 specifically includes: An adaptive dehazing algorithm based on the atmospheric scattering model is used to process images in the battlefield image dataset. Combined with dynamic normalization of illumination compensation, an enhanced battlefield image dataset is obtained by dynamically estimating atmospheric light and transmittance. The formula is as follows: ， in, Indicates the image in the enhanced battlefield image dataset at coordinates The pixel value at ; Indicates the coordinates of the image in the battlefield image dataset The pixel value at ; Indicates the channel adaptive atmospheric light estimation Channel Position pixel value, Represents the coordinates in the transmittance map The transmittance value at express The local illumination compensation coefficient of the position, Represents a constant.

4. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 3 is characterized in that: In step S2, when calculating In the process of dynamic estimation, the smoke concentration of each channel is dynamically estimated by combining the dark channel prior and the exponential weighted average of the local window.

5. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 4 is characterized in that: The 3×3 convolution process of branch 2 in step S3 adopts a dynamic kernel generation mechanism to dynamically adjust the convolution weights according to the input features, so that the model automatically increases the receptive field in dense obstacle areas and focuses on local details in small target areas.

6. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 5, characterized in that: In step S3, the feature fusion module adopts a feature fusion method based on the attention gating mechanism, dynamically weights the contribution of each branch, and calculates channel-level correlation, so that the model relies on the original features of branch 3 when the scene is obscured by smoke, while focusing on the deep features of branches 2 and 1 in clear scenes.

7. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 6, characterized in that: In step S3, the loss function uses contrast loss to enhance intra-class compactness. By adopting the transmittance-aware contrast loss term and dynamically adjusting the contrast loss weight in combination with smoke density and illumination changes, the model can achieve robust feature learning under smoke occlusion and illumination changes, forcing obstacles of the same type to cluster in the feature space and separate heterogeneous objects.

8. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 7, characterized in that: During the loss function calculation process in step S3, in the calculation process of the transmittance perception contrast loss term, the dynamic weighting mechanism of the transmittance and illumination perception contrast loss is used to achieve the weakened constraint of the smoke area and the enhanced separation of the high illumination area.

9. The artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 8, characterized in that: In step S3, the convolutional neural network model uses a transmittance-channel joint gated classifier function to dynamically reconstruct the classification weights by fusing features with the transmittance map and the cross-modal interaction of the illumination coefficient, and finally outputs the category probability, specifically including: Transmittance-aware channel enhancement: A learnable smoke-sensitive channel selection matrix is ​​constructed, and high-frequency compensation weights are calculated through cross-attention between the transmittance map and feature channels. Illumination-robust channel suppression: Designing an illumination-channel covariance matrix to suppress noisy channel responses under high illumination variations, and calculating the correlation between illumination coefficients and feature global statistics through outer products to achieve channel-level dynamic suppression. The enhanced and suppressed features are interacted cross-modally to generate a dynamic classification weight matrix. The transmittance and illumination perception features are fused through bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression to finally output the category probability.

10. An artificial intelligence-based battlefield medical drone obstacle avoidance and recognition device, which executes the artificial intelligence-based battlefield medical drone obstacle avoidance and recognition method according to claim 1, characterized in that: include: A battlefield image acquisition unit is used to acquire battlefield image data and annotate it to obtain a battlefield image data set; Graphics enhancement unit, used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset; A model training unit is used to build a convolutional neural network model and train the convolutional neural network model using images from the enhanced battlefield image dataset. The training process is optimized using a loss function to obtain a trained model. The result generation unit is used to obtain the image to be detected, and input the battlefield image to be detected into the trained model to identify and classify different obstacles in the image.

Images