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

Abstract

The present application relates to a kind of battlefield medical unmanned aerial vehicle obstacle avoidance identification method and equipment based on artificial intelligence, belong to artificial intelligence and data processing technology field.It includes the following steps: obtaining battlefield image data and marking, obtain battlefield image dataset;Image in data set is adaptively normalized and dynamically enhanced, and enhanced battlefield image dataset is obtained;Convolutional neural network model is constructed, model includes 3 different branches and feature fusion module, branch 2 adopts dynamic kernel generation mechanism in convolution process, and feature fusion module introduces attention gate mechanism;Through enhanced battlefield image, convolutional neural network model is trained;Loss function is used to optimize training process, and the trained model is obtained;To be detected battlefield image is input into the trained model, and different obstacles in image are identified and classified.The present application can improve the accuracy and robustness of battlefield medical unmanned aerial vehicle obstacle avoidance identification classification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data processing technology, specifically relating to an obstacle avoidance and recognition method and device for battlefield medical drones based on artificial intelligence. Background Technology

[0002] In modern battlefield environments, medical drones are widely used in emergency rescue and casualty evacuation missions, and their efficiency and timeliness are crucial for the survival of wounded soldiers 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 are diverse, including buildings, wounded soldiers, explosives, smoke, and fire, and these obstacles often appear in extreme lighting conditions, smoke-covered environments, or dynamic environments, requiring drones to possess extremely strong image processing and obstacle avoidance capabilities when performing missions.

[0003] Existing methods combine deep learning algorithms to analyze construction road surface images captured by UAVs to predict the density of rockfill and guide the roller's travel 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 input into the UAV target recognition model based on priority ranking results to improve recognition efficiency and accuracy. Furthermore, wind speed and light intensity are acquired through the UAV's onboard environmental data acquisition module to calculate the number of images to capture, including images of targets such as transformers. Finally, image recognition is performed by an image recognition terminal to adapt to different spatial environments and reduce the recognition error rate. These methods collectively promote the development of UAV obstacle avoidance and recognition technology, improving the intelligence level and application effectiveness of UAVs. However, the following problems still need to be addressed: Most existing technologies use fixed convolution kernels, which are difficult to handle obstacles with huge scale variations in battlefield images, resulting in a fixed perception range of the network and difficulty in adapting to the distribution density and size variations 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; When dealing with environments with drastic changes in lighting or smoke obscuration, traditional methods often rely on global normalization or simple dehazing algorithms, which leads to unstable image quality and affects the effect of feature extraction; When facing complex battlefield scenes, traditional methods often cannot handle various types of obstacles and background interference, resulting in large differences in the performance of the model in different environments. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides an obstacle avoidance and recognition method and device for battlefield medical drones based on artificial intelligence.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention provides an obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence, comprising the following steps:

[0007] S1. Acquire and label battlefield image data to obtain a battlefield image dataset;

[0008] S2. Adaptive illumination normalization and dynamic dehazing enhancement are performed on the images in the dataset to obtain an enhanced battlefield image dataset;

[0009] S3. Construct a convolutional neural network model, which includes three different branches and a feature fusion module. Branch 1 uses a 1×1 convolutional kernel; branch 2 uses a 3×3 convolutional kernel and employs a dynamic kernel generation mechanism during the convolution process; branch 3 uses identity mapping; the feature fusion module introduces an attention gating mechanism; the convolutional neural network model is trained using images from an enhanced battlefield image dataset; the training process is optimized using a loss function to obtain a trained model;

[0010] S4. The battlefield image to be detected is input into the trained model, which identifies and classifies different obstacles in the image.

[0011] Furthermore, in step S1, image data from different battlefield environments is collected. The image data should cover various complex obstacles and environmental conditions, including buildings, wounded, explosion debris, smoke, and fire. The labeled categories include building debris, wounded, explosives, and drones.

[0012] Furthermore, step S2 specifically includes:

[0013] 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. The formula is as follows:

[0014] ,

[0015] in, This indicates that the images in the enhanced battlefield image dataset are in coordinates. Pixel value at; Represents the coordinates of images in the battlefield image dataset. Pixel value at; In the adaptive atmospheric light estimation of the channel, the first... channel Location pixel value, Represents the coordinates in the transmittance graph The transmittance value at that location, express The local illumination compensation coefficient of the location. Represents a constant.

[0016] Furthermore, in calculation During the process, the smoke concentration of each channel is dynamically estimated by combining the prior of the dark channel and the exponential weighted average of the local window.

[0017] Furthermore, in step S3, the 3×3 convolution process of branch 2 adopts a dynamic kernel generation mechanism, which dynamically adjusts 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.

[0018] Furthermore, in step S3, the feature fusion module adopts a feature fusion method based on 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 smoke is obscured, while focusing on the deep features of branch 2 and branch 1 in clear scenes.

[0019] Furthermore, in step S3, the loss function adopts contrastive loss to enhance intra-class compactness. By using a transmittance-aware contrastive loss term and dynamically adjusting the contrastive loss weights in combination with smoke density and illumination changes, the model achieves robust feature learning under smoke occlusion and illumination changes, forcing similar obstacles to cluster in the feature space and separating dissimilar ones.

[0020] Furthermore, in the loss function calculation process of step S3, during the calculation of the transmittance-sensing contrast loss term, the weakening constraint of the smoke region and the strengthening separation of the high-illuminance region are achieved through the dynamic weighting mechanism of the contrast loss of transmittance and illumination perception.

[0021] Furthermore, in step S3, the convolutional neural network model employs a transmissivity-channel joint gated classifier function. Through cross-modal interaction of fusing features with transmissivity maps and illumination coefficients, it dynamically reconstructs classification weights and ultimately outputs class probabilities, specifically including:

[0022] Transmittance sensing channel enhancement: Construct a learnable smoke-sensitive channel selection matrix and calculate high-frequency compensation weights through cross-attention of transmittance map and feature channels;

[0023] Illumination-robust channel suppression: Design an illumination-channel covariance matrix to suppress noise channel responses under high illumination variations. Calculate the correlation between illumination coefficients and feature global statistics through outer product to achieve channel-level dynamic suppression.

[0024] The enhanced and suppressed features are subjected to cross-modal interaction to generate a dynamic classification weight matrix. Transmittance and illumination perception features are fused through bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression, ultimately outputting the class probability.

[0025] This invention also provides an AI-based obstacle avoidance and recognition device for battlefield medical drones, which executes the aforementioned AI-based obstacle avoidance and recognition method for battlefield medical drones, including:

[0026] The battlefield image acquisition unit is used to acquire and annotate battlefield image data to obtain a battlefield image dataset.

[0027] The image enhancement unit is used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset.

[0028] The model training unit is used to build a convolutional neural network model and train the convolutional neural network model by collecting images from an enhanced battlefield image dataset. The training process is optimized using a loss function to obtain a trained model.

[0029] The result generation unit is used to acquire 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.

[0030] The advantages of this invention are:

[0031] This invention employs an adaptive dehazing algorithm based on an atmospheric scattering model, combined with a dynamic normalization method for illumination compensation. This dynamically estimates atmospheric light and transmittance, thus mitigating the impact of smoke interference and illumination variations on image features, significantly enhancing image quality and making subsequent feature extraction more accurate. By designing a multi-branch convolutional neural network model, incorporating convolutional kernels of different sizes and identity mapping, it effectively extracts spatial features and local details of the image. The multi-branch structure can adaptively handle complex obstacles and enhance the network's robustness in varying scenarios. The identity mapping branch preserves original information, avoiding noise interference with key features. Traditional convolutional neural networks typically use fixed convolutional kernels; this invention proposes a dynamic convolutional kernel generation mechanism that dynamically adjusts the kernel weights based on changes in input features, enabling the model to adapt to multi-scale obstacles and improve performance in complex battlefield environments, such as automatically enhancing features in high-density or large-size obstacle regions. Extraction Capabilities: This invention proposes a feature fusion method based on an attention-gated mechanism, which can automatically adjust the feature contribution of different convolutional branches according to the complexity of different scenes, avoiding feature redundancy and noise interference that may occur in traditional feature stitching methods, and optimizing the fusion effect of image features. A transmittance-aware contrast loss is adopted to enhance intra-class compactness and inter-class separation, especially under conditions of smoke occlusion and changing illumination. By dynamically adjusting the weights of the loss function, the recognition accuracy of obstacles is optimized. In addition, the illumination compensation mechanism improves image quality under high illumination conditions by dynamically calculating global and local illumination mean values, enhancing the robustness of the system. The proposed transmittance-channel joint gating classifier function fuses transmittance maps and illumination information with image features through cross-modal interaction, dynamically adjusting classification weights. This enhances information retention in smoke areas and improves inter-class separation in areas of strong illumination, thereby improving classification accuracy and robustness. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0033] Figure 1 This is an architectural diagram of the method of the present invention;

[0034] Figure 2 This is a diagram illustrating the spatial adaptability analysis of the dynamic convolution kernel of the present invention.

[0035] Figure 3 This is an attention-gated feature fusion analysis diagram of the present invention;

[0036] Figure 4 This is a graph showing the performance verification and analysis of transmittance sensing loss of the present invention;

[0037] Figure 5This is a performance verification diagram of the illumination compensation mechanism of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0039] Example 1

[0040] In this embodiment, as Figure 1 As shown, an obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence is provided, and the specific steps include:

[0041] S1. Acquire and label battlefield image data to obtain a battlefield image dataset;

[0042] Specifically, in obstacle avoidance and recognition tasks for battlefield medical drones, it is necessary to collect image data from different battlefield environments. This image data should cover a wide range of scenarios, including various complex obstacles and environmental conditions such as buildings, wounded personnel, explosion debris, smoke, and fire. Because battlefield images typically exhibit high diversity, including varying lighting conditions, obstructions, and dynamic objects, data acquisition must ensure comprehensiveness and diversity, covering a variety of possible battlefield scenarios. During the acquisition process, it is essential to ensure that each image is sufficiently representative, reflecting different battlefield obstacles and their interaction with the surrounding environment.

[0043] Battlefield image data annotation involves accurately labeling every obstacle and key feature in the battlefield images to ensure that each image sample corresponds to a specific label. Annotations typically include the obstacle's category, location, and relative position. Label categories include building debris, wounded personnel, explosives, and drones. Annotation is performed manually by professionals to ensure accuracy and consistency.

[0044] S2. Adaptive illumination normalization and dynamic dehazing enhancement are performed on the images in the dataset to obtain an enhanced battlefield image dataset;

[0045] Specifically, battlefield image data is characterized by drastic changes in lighting, smoke obscuration, and dynamic motion blur. At the same time, battlefield images are usually affected by smoke obscuration and changes in lighting, leading to image feature degradation.

[0046] Traditional normalization only linearly scales pixel values, which cannot eliminate the interference of smoke and motion blur on texture features;

[0047] An adaptive dehazing algorithm based on an atmospheric scattering model is used to process images in a battlefield image dataset. Combined with dynamic normalization for illumination compensation, an enhanced battlefield image dataset is obtained by dynamically estimating atmospheric light and transmittance, thus solving the feature degradation problem under smoke interference. The formula is as follows:

[0048] ,

[0049] in, This indicates that the images in the enhanced battlefield image dataset are in coordinates. Pixel value at; Represents the coordinates of images in the battlefield image dataset. Pixel value at; In the adaptive atmospheric light estimation of the channel, the first... channel Location pixel value, express The local illumination compensation coefficient of the location. This represents a constant used to avoid division by zero errors; it is set to 0.0001.

[0050] Specifically, in calculation During the process, the smoke concentration of each channel is dynamically estimated by combining the dark channel prior and using the exponential weighted average of local windows. The calculation formula is expressed as follows:

[0051] ,

[0052] It should be noted that battlefield smoke distribution is uneven; for example, the distribution of dense smoke is uneven in the localized area of ​​an explosion. (The exponential function term...) To suppress interference from non-smoke areas and enhance the estimation accuracy of smoke areas; among which, This represents a local window, such as a region defined in an image. Its size can be set according to specific application requirements, such as 3×3 or 5x5. This window is used to calculate the statistical characteristics of local pixel values, which helps to deduce the local brightness and transmittance of the image. Represents the pixel coordinates within a local window. This represents an exponential function with the natural constant as its base. Represents the L2 norm; The dark channel prior is represented by the following formula:

[0053] ,

[0054] in, This represents the minimum value calculation for all pixel coordinates within a local window. Represents the image in the battlefield image dataset. channel Location pixel value; This represents the dark channel smoothing coefficient, set to... ; Represents the coordinates in the transmittance graph The transmittance value at that location is expressed by the following formula:

[0055] ,

[0056] in, This refers to the smoke density adjustment factor. Smoke density varies greatly on the battlefield. Adjusting this factor balances defogging and detail retention. It is set to... ; This represents the minimum value in the red, green, and blue channels, used to represent the minimum color channel information at that point, in order to estimate the dark channel of the image and adapt to the non-uniform attenuation of the RGB channels by smoke in battlefield images, such as the phenomenon that the red channel is easily affected by firelight. Represents the image in the battlefield image dataset. channel Position pixel value, number The channel can be red, green, or blue; by calculating the ratio of the global mean to the local mean, drastic changes in battlefield lighting are dynamically compensated, as expressed by the following formula:

[0057] ,

[0058] in, Represents the global pixel mean. This represents the mean of the sliding window; This represents a constant used to avoid division by zero errors; it is set to 0.0001.

[0059] S3. Construct a convolutional neural network model, which includes three different branches and a feature fusion module. Branch 1 uses a 1×1 convolutional kernel; branch 2 uses a 3×3 convolutional kernel and employs a dynamic kernel generation mechanism during the convolution process; branch 3 uses identity mapping; the feature fusion module introduces an attention gating mechanism; the convolutional neural network model is trained using images from an enhanced battlefield image dataset; the training process is optimized using a loss function to obtain a trained model;

[0060] Branch 1 uses 1×1 convolutional kernels to change the number of channels in the feature map, reducing dimensionality or increasing network depth while preserving spatial information. Battlefield images may contain a large amount of redundant information. When the data volume is large, 1×1 convolution can effectively reduce dimensionality 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 one 1×1 convolutional kernel. The convolution operation of multiple 1×1 convolutional 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 1024x1024 feature map is obtained through one 1×1 convolutional kernel.

[0061] Traditional multi-branch networks use fixed convolutional kernels, which are difficult to adapt to the multi-scale characteristics of battlefield obstacles, such as the size differences of explosion debris, wounded limbs, and building debris. Branch 2 determines the convolutional kernel weight parameters through dynamic kernel weights and uses padding to maintain the spatial dimension of the feature map and extract spatial features. In battlefield images, complex obstacles, such as buildings and wounded soldiers, have large scale variations. Dynamic convolutional kernels can adaptively handle obstacles of different sizes, enhancing adaptability to different types of targets. Specifically, the 3×3 convolution in Branch 2 uses a dynamic kernel generation mechanism, dynamically adjusting the convolutional weights according to the input features. This allows the model to automatically increase the receptive field in dense obstacle areas and focus on local details in small target areas. The formula is as follows:

[0062] ,

[0063] ,

[0064] in, Indicates branch 2 The initial convolutional kernel weights of the layer; This represents the dynamic convolutional kernel weights of the k-th layer in branch 2. The convolutional kernel weights are dynamically adjusted according to changes in the input features to adapt to changes in different scenarios. Indicates branch 2 The kernel weight increment of the layer is dynamically adjusted based on the input features during training. The dynamic calculation of the kernel weight can adaptively adjust the kernel according to the input features, which solves the limitations of traditional fixed kernels when dealing with multi-scale obstacles. For example, when the input is a smoke-covered area, the global average pooling operation outputs low activation values, and the fully connected network reduces the kernel weight to reduce noise propagation. This represents the global average pooling operation, which reduces the dimensionality of the feature map by calculating the average value of the entire feature map, thereby obtaining features with more global information. Let be the input feature map of the k-th layer in branch 2, representing the feature representation before the convolution operation in that layer; It is a fully connected network with the LeakyReLU activation function. By processing the features after global average pooling, the network can automatically adjust the size of the receptive field according to different battlefield environments, thereby focusing on details in dense areas and maintaining broad perception in sparse areas.

[0065] Branch 3 is identity mapping, where the input feature map is directly passed through the network unchanged to preserve original information and mitigate the vanishing gradient problem. In battlefield images, some features may be very weak and easily lost due to noise and occlusion. Identity mapping provides direct transmission of original information, ensuring that the network does not lose critical information when processing complex images.

[0066] In battlefield images, smoke and occlusion can lead to the loss of information in some areas. Conventional methods often directly concatenate the outputs of the three branches, which can easily result in feature redundancy and sensitivity to noise. Specifically, the feature fusion module adopts a feature fusion method based on an attention gating mechanism, dynamically weighting the contributions of each branch. Through channel-level correlation calculation, the model relies on the original features of branch 3 when there is smoke or occlusion, while focusing on the deeper features of branch 2 and branch 1 in clear scenes. The formula is expressed as follows:

[0067] ,

[0068] ,

[0069] in, This represents the Sigmoid activation function; This represents a compressed-excitation module consisting of two 1×1 convolutional layers, which maps the channel-concatenated features to a three-dimensional attention weight space. For the first The attention weights of the branches are dynamically adjusted by the model through the attention gating mechanism, so as to better integrate features of different scales and levels. The feature map output for branch 1; The feature map output for branch 2; The feature map output for branch 3; Indicates to , and Perform channel splicing operation; This results in a multi-scale feature map after fusion. Through dynamic weighting, it ensures that more original information is relied upon when there is severe occlusion, while deeper features are relied upon when the image is clear. The attention weights for branch 1; The attention weights for branch 2; The attention weights for branch 3; This indicates multiplication by channel.

[0070] Specifically, during model training, conventional single cross-entropy loss cannot distinguish similar obstacle categories; for example, wounded soldiers and sandbags are easily considered similar obstacles. Contrast loss is used to enhance intra-class compactness. By employing a transmittance-aware contrast loss term and dynamically adjusting the contrast loss weights based on smoke density and lighting changes, the model achieves robust feature learning under smoke occlusion and lighting variations, forcing similar obstacles to cluster in the feature space and separating out-of-class obstacles. The formula is as follows:

[0071] ,

[0072] in, To improve the contrast loss function, the model's ability to identify obstacles in multi-obstacle environments is enhanced by strengthening intra-class compactness and inter-class separation. For example, in smoke areas, transmittance-aware contrast loss helps distinguish the features of different obstacles. The number of training samples in a single batch;

[0073] In order to be with the first A set of positive samples of the same type; The feature vector extracted from the i-th sample by a multi-branch convolutional neural network; Let i be the battlefield image sample input into the multi-branch convolutional neural network; This is the feature vector extracted from the p-th sample (of the same type) by a multi-branch convolutional neural network; Let p be the battlefield image sample input to the multi-branch convolutional neural network, and let be the feature vector of the same type of sample. The feature vector extracted from the nth sample (outlier sample) by a multi-branch convolutional neural network; The nth battlefield image sample input into the multi-branch convolutional neural network is a feature vector of outlier samples; In order to be with the first A set of negative samples from different classes; This is the balance coefficient; This indicates that feature space aggregation of similar samples is achieved through feature distance constraints; To maximize the function, negative values ​​are truncated to zero, ensuring that the contrastive loss only generates gradient updates when the feature distance between samples of the same class is greater than the feature distance between samples of different classes. This represents the transmittance-sensitized contrast loss term.

[0074] Specifically, to address the issues of feature blurring caused by smoke obscuration and inter-class confusion under drastic changes in illumination, a dynamic weighting mechanism of transmittance and illumination-perceived contrast loss is used in the calculation of the transmittance-perceived contrast loss term to achieve weakened constraints in smoke regions and enhanced separation in high-illumination regions. The formula for the transmittance-perceived contrast loss term is as follows:

[0075] ,

[0076] in, This represents the mean value of the transmittance map; The weakening contrast term for the smoke area is represented by the calculation method as follows: When the smoke is dense, that is, when the transmittance is low, the feature distance constraint of similar samples is reduced to avoid forcing the model to learn invalid features under the smoke occlusion. For the sample The mean transmittance of the corresponding region represents the weighted average transmittance of the positive sample comparison term, indicating the smoke-covered area. Reduce feature clustering constraints, i.e. , for Local area; The coordinates in the transmittance graph The transmittance value at that location is calculated as follows: ; For the battlefield image dataset, the first image in the dataset... channel Position pixel value, number The channel can be red, green, or blue; For the first channel adaptive atmospheric light estimation channel The location pixel value represents the adaptive atmospheric light value for each channel; Indicates the cardinality of the set of positive samples; The enhancement term for light-sensitive negative samples is calculated as follows: When the illumination changes drastically, that is, when the local illumination compensation coefficient is large, the separation intensity of heterogeneous samples is enhanced, and inter-class confusion in the bright areas is prevented, such as preventing inter-class confusion between firelight and orange obstacles. For the sample The average illumination compensation coefficient, Enhance inter-class separation intensity in regions with high light intensity variations; for The local illumination compensation coefficient for the location; Indicates the cardinality of the negative sample set; The variance of the transmittance plot; The global mean of the transmittance map is calculated as follows: , and These represent the height and width of the transmittance map, respectively. The ratio of the variance to the mean of the transmittance map is used to force the model to automatically balance the loss weights when the smoke distribution is uneven.

[0077] Specifically, the convolutional neural network model employs a transmissivity-channel joint gated classifier function. By fusing cross-modal interactions between features, transmissivity maps, and illumination coefficients, it dynamically reconstructs classification weights and ultimately outputs class probabilities, including the following steps:

[0078] (1) Transmittance sensing channel enhancement: A learnable smoke-sensitive channel selection matrix is ​​constructed, and high-frequency compensation weights are calculated through cross-attention of the transmittance map and feature channels. The formula is as follows:

[0079] ,

[0080] in, Enhance the attention weight of the channel, with the dimension being It is used to select smoke-sensitive feature channels. By cross-attention between the transmittance map and the feature channels, it compensates for the high-frequency information attenuation caused by smoke. That is, the channel weight corresponding to the smoke region with low transmittance is reduced, and the response of the noise channel is reduced. The Softmax function performs normalization along the channel dimension. This is the fused multi-scale feature map; for transpose; This is a global max pooling operation; This is a global average pooling operation; A learnable spatial transmittance adaptation matrix; This represents the number of channels in the fused multi-scale feature map.

[0081] (2) Illumination-robust channel suppression: An illumination-channel covariance matrix is ​​designed to suppress the noise channel response under high illumination variation. The correlation between the illumination coefficient and the characteristic global statistics is calculated by outer product to achieve channel-level dynamic suppression. The formula is as follows:

[0082] ,

[0083] in, This represents the channel inhibition coefficient. Represents the tensor outer product. This represents the local illumination compensation coefficient. Indicates will and By performing an outer product, covariance features are generated, which suppress channel noise caused by drastic changes in illumination, such as channel noise caused by overexposed areas under strong light.

[0084] (3) The enhanced and suppressed features are subjected to cross-modal interaction to generate a dynamic classification weight matrix. The transmittance and illumination perception features are fused by bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression. The final output class probability is expressed by the following formula:

[0085] ,

[0086] ,

[0087] ,

[0088] ,

[0089] in, This is a feature map after transmittance enhancement; This is a feature map after light suppression; Indicates a channel splicing operation; The dynamically generated classification weight matrix consists of learnable parameters, which are used to map the global statistics of transmittance and illumination to the classification space through a multilayer perceptron. for transpose; The classification bias is a learnable parameter. For the final category probability distribution, express Various obstacle categories, such as: wounded personnel, building debris, drones, etc.; This refers to the Softmax function.

[0090] S4. The battlefield image to be detected is input into the trained model, which identifies and classifies different obstacles in the image.

[0091] Example 2

[0092] This embodiment performs dynamic convolution kernel spatial adaptability analysis. By constructing a three-dimensional spatial surface, it explores the influence of obstacle distribution density and size on detection accuracy. Figure 2 As shown, compared with traditional fixed convolution kernels Figure 2 (a) Dynamic convolution kernel of this technology Figure 2(b) Adaptability to different battlefield environments: Experimental results show that when the density and size of obstacles in the battlefield environment are high, the average accuracy surface of the traditional method is significantly concave, indicating that its fixed receptive field is difficult to adapt to the complex and ever-changing obstacle distribution. However, the surface generated by this technology remains stable, and the color mapping 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 increment, automatically enhance the feature extraction capability in dense and large obstacle scenes, and overcome the limitations of traditional methods in complex battlefield environments.

[0093] Example 3

[0094] This embodiment reveals the working principle of the multi-branch feature fusion mechanism by comparing the branch weight allocation in clear scenes and smoke-obscured scenes. The stacked bar chart corresponding to the attention-gated feature fusion analysis is shown below. Figure 3 As shown, Figure 3 (a) shows the attention weight distribution for a clear scene. Figure 3 (b) For the attention weight distribution under smoke occlusion, in a smoke-filled environment, the weight of the third branch, which represents the preservation of original information, is significantly increased, while in a clear scene, the dynamic convolution branch dominates. The dynamic adjustment is due to the attention gating mechanism's real-time perception of the battlefield environment features. When the image quality deteriorates due to smoke, the system automatically enhances the contribution of the identity mapping branch to ensure that the information of key obstacles is not drowned out by noise. The experimental results intuitively demonstrate the intelligent feature selection capability achieved by this technology through channel-level weight allocation, which solves the problem of low information utilization in occluded scenes in traditional methods.

[0095] Example 4

[0096] like Figure 4 As shown in the figure, this embodiment verifies the effectiveness of transmittance-sensory contrast loss by analyzing the impact curve of smoke concentration changes on obstacle recall. The experimental curves show that the recall rate of the traditional method decreases sharply with increasing smoke concentration, and the confidence interval gradually expands. In contrast, the curve generated by this technology decreases more smoothly with smaller fluctuations, and its performance stability in high-concentration areas is significantly better than that of the comparative method. This is because this technology automatically reduces invalid feature constraints in areas with uneven smoke distribution through a transmittance variance adjustment mechanism, while strengthening the inter-class discrimination of valid 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.

[0097] Example 5

[0098] This embodiment uses a dual-axis graph to comprehensively evaluate the impact of changes in light intensity on the detection system. Figure 5(a) shows that the accuracy of the traditional method decreases parabolically with increasing light intensity, while the curve of this technique only shows slight fluctuations under extreme light conditions; Figure 5 (b) The error distribution box plot further shows that the error dispersion of this technology is reduced by more than 50% compared with the traditional method, indicating that the illumination compensation coefficient dynamically corrects the local brightness difference. By calculating the ratio of the global and local illumination mean, the pixel distortion in the overexposed area is effectively suppressed. The performance comparison on the logarithmic 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 battlefield illumination conditions.

[0099] Example 6

[0100] This embodiment provides an AI-based obstacle avoidance and recognition device for battlefield medical drones, which executes the AI-based obstacle avoidance and recognition method for battlefield medical drones described in Embodiment 1, including:

[0101] The battlefield image acquisition unit is used to acquire and annotate battlefield image data to obtain a battlefield image dataset.

[0102] The image enhancement unit is used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset.

[0103] The model training unit is used to build a convolutional neural network model and train the convolutional neural network model by collecting images from an enhanced battlefield image dataset. The training process is optimized using a loss function to obtain a trained model.

[0104] The result generation unit is used to acquire 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.

[0105] 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 foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence, characterized in that, Includes the following steps: S1. Acquire and label battlefield image data to obtain a battlefield image dataset; S2. Adaptive illumination normalization and dynamic dehazing enhancement are performed on the images in the dataset to obtain an enhanced battlefield image dataset; S3. Construct a convolutional neural network model, which includes three different branches and a feature fusion module. Branch 1 uses a 1×1 convolutional kernel; branch 2 uses a 3×3 convolutional kernel and employs a dynamic kernel generation mechanism during convolution; branch 3 uses identity mapping, which directly transmits the input feature map through the network unchanged, preserving the original information; the feature fusion module introduces an attention gating mechanism; train the convolutional neural network model using images from an enhanced battlefield image dataset; optimize the training process using a loss function to obtain a trained model; specific content includes: The feature fusion module employs an attention-gating-based feature fusion method, dynamically weighting the contributions of each branch. Through channel-level correlation calculation, the model relies on the original features of branch 3 when obscured by smoke, while focusing on the deeper features of branches 2 and 1 in clear scenes. The formula is expressed as follows: ， ， in, This represents the Sigmoid activation function; This represents a compressed-excitation module consisting of two 1×1 convolutional layers; For the first Attention weights for branches; The feature map output for branch 1; The feature map output for branch 2; The feature map output for branch 3; Indicates to , and Perform channel splicing operation; This is the fused multi-scale feature map; The attention weights for branch 1; The attention weights for branch 2; The attention weights for branch 3; This indicates multiplication by channel; The convolutional neural network model employs a transmissivity-channel joint gated classifier function. By fusing cross-modal interactions between features, transmissivity maps, and illumination coefficients, it dynamically reconstructs classification weights and ultimately outputs class probabilities, specifically including: Transmittance-sensing channel enhancement: A learnable smoke-sensitive channel selection matrix is ​​constructed. High-frequency compensation weights are calculated across the transmittance map and feature channels to obtain the channel enhancement attention weights; the formula is as follows: ， in, Increase the attention weight of the channel; The Softmax function performs normalization along the channel dimension. This is the fused multi-scale feature map; for Transpose of; This is a global max pooling operation; This is a global average pooling operation; A learnable spatial transmittance adaptation matrix; This represents the number of channels in the fused multi-scale feature map. Represents a transmittance diagram; Illumination-robust channel suppression: An illumination-channel covariance matrix is ​​designed to suppress noise channel responses under high illumination variations. The correlation between the illumination coefficient and characteristic global statistics is calculated through outer product to achieve channel-level dynamic suppression; the formula is expressed as follows: ， in, This represents the channel inhibition coefficient. Represents the tensor outer product. Indicates the local illumination compensation coefficient; The enhanced and suppressed features are subjected to cross-modal interaction to generate a dynamic classification weight matrix. Transmittance and illumination-sensing features are fused through bilinear pooling, and a gating mechanism is used to balance high-frequency compensation and noise suppression, ultimately outputting the class probability; the formula is expressed as follows: ， ， ， ， in, This is a feature map after transmittance enhancement; This indicates a channel splicing operation; This is a feature map after light suppression; The classification weight matrix is ​​dynamically generated; for Transpose of; For classification bias; For the final category probability distribution, express Types of obstacles; For the Softmax function; The loss function employs contrastive loss to enhance intra-class compactness. By using a transmittance-aware contrastive loss term and dynamically adjusting the contrastive loss weights in conjunction with smoke density and illumination changes, the model achieves robust feature learning under smoke occlusion and illumination changes, forcing similar obstacles to cluster in the feature space and separating dissimilar objects. The formula is expressed as follows: ， in, To compare loss functions; The number of training samples in a single batch; In order to be with the first A set of positive samples of the same type; The feature vector extracted from the i-th sample by a multi-branch convolutional neural network; Let i be the battlefield image sample input into the multi-branch convolutional neural network; For the p-th sample, i.e., a sample of the same type, the feature vector extracted by a multi-branch convolutional neural network; The p-th battlefield image sample is input into a multi-branch convolutional neural network; The feature vector extracted from the nth sample by a multi-branch convolutional neural network; Let be the nth battlefield image sample input into the multi-branch convolutional neural network, and be the feature vector of the outlier sample. In order to be with the first A set of negative samples from different classes; This is the balance coefficient; This indicates that feature space aggregation of similar samples is achieved through feature distance constraints; This is a function to find the maximum value. For transmittance-sensitized contrast loss; S4. The battlefield image to be detected is input into the trained model, which identifies and classifies different obstacles in the image.

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

3. The obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence according to claim 2, characterized in that, Step S2 specifically includes: 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. The formula is as follows: ， in, This indicates that the images in the enhanced battlefield image dataset are in coordinates. Pixel value at; Represents the coordinates of images in the battlefield image dataset. Pixel value at; In the adaptive atmospheric light estimation of the channel, the first... channel Location pixel value, Represents the coordinates in the transmittance graph The transmittance value at that location, express The local illumination compensation coefficient of the location. Represents a constant.

4. The obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence according to claim 3, characterized in that, In step S2, during the calculation During the process, the smoke concentration of each channel is dynamically estimated by combining the prior of the dark channel and the exponential weighted average of the local window.

5. The obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence according to claim 4, characterized in that, In step S3, the 3×3 convolution process in branch 2 adopts a dynamic kernel generation mechanism, which dynamically adjusts the convolution weights according to the input features, so that the model can automatically increase the receptive field in dense obstacle areas and focus on local details in small target areas.

6. The obstacle avoidance and recognition method for battlefield medical drones based on artificial intelligence according to claim 5, characterized in that, In step S3, during the calculation of the loss function, the dynamic weighting mechanism of transmittance and illumination-sensed contrast loss is used to separate the weakening constraint of the smoke region from the strengthening of the high-illumination region during the calculation of the transmittance-sensed contrast loss term. The formula for the transmittance-sensed contrast loss term is as follows: ， in, This indicates a weakened contrast in the smoke area; This indicates a light-sensitive negative sample enhancement term; Represents the variance of the transmittance map; This represents the global mean of the transmittance map.

7. An AI-based obstacle avoidance and recognition device for battlefield medical drones, executing the AI-based obstacle avoidance and recognition method for battlefield medical drones as described in claim 1, characterized in that, include: The battlefield image acquisition unit is used to acquire and annotate battlefield image data to obtain a battlefield image dataset. The image enhancement unit is used to perform adaptive illumination normalization and dynamic dehazing enhancement on the images in the dataset to obtain an enhanced battlefield image dataset. The model training unit is used to build a convolutional neural network model and train the convolutional neural network model by collecting images from an 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 acquire 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