An online carbon semantic map construction method based on sparse fusion

Abstract

The application discloses an online carbon semantic map construction method based on sparse fusion, and the method comprises the following steps: 1) obtaining semantic labels of city images based on a target detection neural network; 2) completing sparse map registration of the semantic labels through feature matching, then performing clustering segmentation and plane reconstruction on sparse semantic map points; 3) realizing reconstruction and perception of a semantic scene based on a local plane rolling optimization method of a space-time geometric constraint; and 4) combining carbon emission monitoring point cloud information to construct a carbon semantic map. The application has the advantages that: based on a sparse depth map, a fast carbon semantic map reconstruction method is proposed, reconstruction and perception of a carbon semantic scene are realized, and safe navigation of a UAV performing a complex carbon monitoring task is ensured. Through a large number of outdoor scene reconstruction experiments, it is proved that compared with similar methods, the method can improve the semantic plane reconstruction accuracy by 80.5%, and the calculation speed is improved by 81 times. Compared with other scene perception algorithms based on a dense depth map, the method has lower calculation complexity and lower scale error, and is suitable for an airborne processing platform for complex UAV flight tasks.

Description

Technical Field

[0001] This invention relates to the field of environmental map construction technology, and in particular to a map construction method based on carbon semantic information. Background Technology

[0002] Semantic mapping is a crucial tool for scene understanding and interaction, and several systems have successfully demonstrated the feasibility of semantic mapping using monocular cameras. Tateno et al. proposed a neural network model to simultaneously perform depth estimation and semantic segmentation, fusing depth estimation results from high-gradient regions of LSD-SLAM with those from low-gradient regions of the neural network to improve overall depth estimation accuracy. Dense semantic maps can be obtained by fusing pixel-level semantic segmentation results with dense depth maps using a global segmentation model. Similarly, a method called SemanticFusion uses CNNs for semantic segmentation, RGB-DSLAM for depth mapping, and then combines semantic labels and depth maps using incremental semantic fusion and fully connected stochastic conditional field methods to generate dense semantic maps. Many other semantic SLAM systems are developed for different application scenarios, including indoor augmented reality interaction, monocular SLAM system enhancement, and indoor and outdoor navigation.

[0003] Image semantic information extraction and classification are key issues in the field of computer vision. With the continuous development of digital media applications, extracting semantic information from complex multimedia data and achieving effective understanding of image content has become increasingly important. In the past few years, numerous research efforts have made progress in core areas such as visual feature description, object model construction, and strongly supervised learning.

[0004] In the field of carbon environmental monitoring, the development of numerous monitoring technologies has greatly facilitated environmental data extraction. However, the carbon environmental monitoring field targeted by this invention still faces the following challenges:

[0005] 1) Environmental monitoring information is vague, making it impossible to clearly define the carbon emissions around each area and building;

[0006] 2) The algorithm has high computational complexity and cannot be adapted to scenarios involving handheld, fast-moving devices or large-scale monitoring. Summary of the Invention

[0007] The present invention aims to overcome the above-mentioned shortcomings of the prior art and provide an online carbon semantic map construction method based on sparse fusion.

[0008] A method for constructing an online carbon semantic map based on sparse fusion includes the following steps:

[0009] 1) Obtain semantic labels for city images based on object detection neural networks;

[0010] 2) Sparse map registration of semantic labels is completed through feature matching, and then clustering, segmentation and planar reconstruction of sparse semantic map points are performed;

[0011] 3) A local planar rolling optimization method based on spatiotemporal geometric constraints is used to reconstruct and perceive semantic scenes;

[0012] 4) Construct a carbon semantic map by combining carbon emission monitoring point cloud information;

[0013] Invention flowchart as follows Figure 1 As shown.

[0014] Furthermore, in step 1), the object detection network first rescales the image to a uniform resolution input image, then feeds the image into the neural network to predict the location of the target, the confidence score of the bounding box containing the target, and the conditional probability of classification. Finally, non-maximum suppression is applied to the results to obtain the classification result with the best probability. The workflow of the object detection network is as follows: Figure 2 As shown.

[0015] The network implementation process is as follows: The input image is segmented into S×S grids. Each grid independently predicts B bounding boxes and a confidence score for each box. The confidence score includes the probability that an object exists in the predicted bounding box and the accuracy of the bounding box's location. Furthermore, each grid also needs to predict a conditional probability C, which is the probability that an object belongs to a certain class given that an object exists in that grid. Therefore, the network output is an S×S×(B*5+C) dimensional vector. During prediction, the classification information of each grid and its confidence score are multiplied to obtain the grid's classification-spatial confidence score. Finally, non-maximum suppression is used to find the bounding box and class of each grid that contains the most accurate object location information.

[0016] Semantic labels for city images are obtained based on object detection neural networks. The specific network is as follows: Figure 2 As shown, this network employs 24 convolutional layers and 2 fully connected layers. It eliminates the need for a region proposal network to find the target; direct regression is sufficient for target detection and recognition. The convolutional layers extract image features, while the fully connected layers predict the target location and category probability. This network borrows from the GoogLeNet classification network structure, but unlike GoogLeNet, it uses 1×1 convolutional layers followed by 3×3 convolutional layers to replace the Inception module, reducing computational dimensionality and normalizing the prediction results. The last layer uses a linear activation function, and the other layers use the Rectified Linear Unit (RELU) from equation (1) as the activation function.

[0017]

[0018] Furthermore, this network uses mean squared error as the loss function to optimize the model parameters, that is, the mean squared error between the S×S×(B*5+C) dimensional vector output by the network and the S×S×(B*5+C) dimensional vector corresponding to the real image. The loss function is shown in Equation (2).

[0019]

[0020] Here, coordError, iouError, and classError represent the coordinate error, intersection-over-union (IoU) error, and classification error between the predicted and calibrated data, respectively.

[0021] Furthermore, in step 2), sparse map registration of semantic labels is completed through feature matching, followed by clustering and segmentation of sparse semantic map points and planar reconstruction. Semantic labels are applied to frequently occurring and relatively stable targets in the scene. In the image, the identified region is represented as Equation (3).

[0022] Region c ={L,{x LT ,h,w}} (3)

[0023] Where c represents x in the image plane LT The coordinates of the top left corner of the semantic region are represented by h and w, which represent the height and width of the semantic region, respectively. L represents the semantic label of this region.

[0024] Accordingly, the three-dimensional semantic map points are represented as Equation (4).

[0025] X = {L, x} (4)

[0026] Where X represents a 3D semantic map point in the world coordinate system, x represents a map depth point, and L represents a semantic label. To obtain X, the feature ID obtained from the feature tracker is used as an anchor point between the image and the sparse depth feature points to determine the 3D semantic point X.

[0027] Semantic annotation projects the semantic points in each image into a 3D space. For a plane in 3D space, the feature points in one image are far from sufficient. Therefore, these semantic points need to be incrementally updated, and the point cloud of these spaces is divided into different planes for planar reconstruction.

[0028] The supervoxel method is used to reduce the number of 3D regions. The supervoxel processing procedure is as follows: an adjacency graph of the voxel cloud is constructed, spatial seeds are generated and filtered, clustering is performed using feature distances, and spatial connectivity between clusters is enhanced through iterative clustering algorithms. This process is repeated on the 3D point cloud at a fixed frequency and with a sliding window to obtain supervoxel clusters for subsequent LCCP segmentation.

[0029] The LCCP method is used to solve the point cloud block segmentation problem based on the concavity-convexity relationship between two adjacent planes. We represent the supervoxels obtained from clustering as Equation (5).

[0030] Region w ={L, N, P center {x0, x1, ..., x} N}} (5)

[0031] Where w represents the world coordinate system, L represents the semantic label of this region, N represents the number of hypervoxels, and P represents the number of hypervoxels. center Let {x0, x1, ..., x} represent the coordinates of the center point of this supervoxel. N} represents all three-dimensional coordinate points within the hypervoxel.

[0032] Subsequently, the Random Sampling Consensus (RANSAC) method is used to compute the regression plane of the 3D semantic supervoxels. The RANSAC process involves randomly selecting 3D semantic feature points, estimating the model, and calculating the number of interior points belonging to this model. This process is repeated iteratively a certain number of times, outputting the plane parameters and the indices of the interior points belonging to the plane. This removes outliers and obtains an accurate semantic plane that satisfies the Region... w The semantic plane is denoted by π. These semantic planes, determined by sparse semantic feature points, constitute the preliminary semantic space as shown in equation (6).

[0033] W′={(L1, π′1), (L2, π′2),…(L n ,π′ n (6)

[0034] Where n is the number of semantic planes.

[0035] Furthermore, in step 3), in the reconstruction of the semantic plane with sparse features, the orientation of the semantic plane is one of the most difficult parameters to recover accurately, especially for small targets containing a few feature points. However, as the number of semantic planes accumulated in the space increases, additional information can be used to locally optimize the semantic planes in the space. Therefore, a quality loss function for the semantic plane is established, and a local optimization method for the semantic map is proposed around this loss function.

[0036] A local planar rolling optimization method based on spatiotemporal geometric constraints achieves semantic scene reconstruction and perception: such as Figure 2 As shown, the semantic plane of the input is defined as π. k The number of points falling in the plane is The center coordinates of the plane are The normal vector of the plane is The coefficients of the plane are (A, B, C, D).

[0037] Then, the generated plane is added to the global semantic map. This section divides this step into two sub-steps: insertion and update. Insertion step: In the local plane, using the camera's heading and the normal vectors of other planes, a plane quality loss function with respect to orientation error is established.

[0038]

[0039] in It is the normal vector of all other local planes. It is an index of the local plane. It is the heading vector formed by the coordinates of the camera's head and tail within the sliding window: (like Figure 3 (As shown).

[0040] Update steps: Try to find the plane index closest to the input plane. This process can be represented as shown in equation (8).

[0041]

[0042] Where t is the index number of the plane closest to the newly added plane. If the newly added plane is π... k The number of interior points contained is greater than that of the historical plane π. k Then use π k instead of π t Accordingly, the planar mass loss function for the update step is also defined as equation (9).

[0043]

[0044] in Indicates the proportion of additional interior points. This represents the probability of successfully updating the plane.

[0045] The sum of each error term in the insertion and update is minimized using an unconstrained optimization formula, as shown in Equation (10), to obtain the optimal plane parameters.

[0046]

[0047] The optimization function described above does not contain differential terms, so a direct method is used to solve the function optimization problem. Optimization mainly involves two steps: detection search and pattern shifting. The purpose of detection search is to find a salient point around the initial search point (reference point) to determine a favorable direction. Pattern shifting involves moving the salient point to the new reference point. These two steps are repeated alternately, gradually moving the function to its minimum.

[0048] Finally, all optimized semantic planes were obtained, and the final semantic space was expressed as Equation (11).

[0049] W={(L1, π1), (L2, π2),…(L n , π n (11)

[0050] Furthermore, in step 4), a differential optical depth (DOD) method is used to obtain the concentration of carbon emissions in space. The quantification index DOD conforms to the following standard, where F is the best-fit transmittance, used to determine the CO2 path density (unit: m^(-2)):

[0051]

[0052] Or convert to path average mixing ratio Unit: (ppm), where r is the target distance measured by the rangefinder, Δσ is the differential absorption cross section of carbon dioxide, and N air It is the air density:

[0053]

[0054] For each DOD value, the model covariance matrix and the residual standard deviation (root mean square error) of the fitted data are proportionally transformed to provide a meaningful confidence estimate. The UAV monitors data at four arbitrary points during flight, acquires the distances to each point, calculates the DOD for each point, and plots the relevant curves. Finally, the concentration of the relevant carbon-containing gas is derived from the DOD and the following formula:

[0055]

[0056] This invention provides an online carbon semantic map construction method based on sparse fusion. By incorporating scale observations from inertial sensors into monocular visual odometry through sensor pre-integration and tightly coupled sensor observation optimization methods, it provides high-precision attitude estimation and sparse depth map construction. Key semantic targets in the scene are extracted using target-level semantic detection methods. Sparse map registration of semantic labels is completed through feature matching, followed by clustering and segmentation of sparse semantic map points and planar reconstruction. Finally, semantic scene reconstruction and perception are achieved through a local planar rolling optimization method based on spatiotemporal geometric constraints.

[0057] The advantages of this invention are: based on sparse depth maps, a fast carbon semantic map reconstruction method is proposed, realizing the reconstruction and perception of carbon semantic scenes, ensuring safe navigation for UAVs performing complex carbon monitoring tasks. Extensive outdoor scene reconstruction experiments demonstrate that this method improves semantic plane reconstruction accuracy by 80.5% compared to similar methods, while simultaneously increasing computational speed by 81 times. Compared to other scene perception algorithms based on dense depth maps, this method has lower computational complexity and lower scale error, making it suitable for airborne processing platforms carrying out complex UAV flight missions. Attached Figure Description

[0058] Figure 1 This is a flowchart of the present invention.

[0059] Figure 2 This is a flowchart of the target detection network in this invention.

[0060] Figure 3 It is a spherical subspace graph of the semantic plane and normal vector, camera trajectory, camera direction vector, and semantic plane normal vector in this invention. Detailed Implementation

[0061] The specific implementation process of the present invention will be described below with reference to the accompanying drawings.

[0062] A method for constructing an online carbon semantic map based on sparse fusion includes the following steps:

[0063] 1) Obtain semantic labels for city images based on object detection neural networks, such as... Figure 2As shown, this network employs 24 convolutional layers and 2 fully connected layers. It eliminates the need for a region proposal network to find the target; direct regression is sufficient for target detection and recognition. The convolutional layers extract image features, while the fully connected layers predict the target location and category probability. This network borrows from the GoogLeNet classification network structure, but unlike GoogLeNet, it uses 1×1 convolutional layers followed by 3×3 convolutional layers to replace the Inception module, reducing computational dimensionality and normalizing the prediction results. The last layer uses a linear activation function, and the other layers use the Rectified Linear Unit (RELU) from equation (4) as the activation function.

[0064]

[0065] Furthermore, this network uses mean squared error as the loss function to optimize the model parameters, that is, the mean squared error between the S×S×(B*5+C) dimensional vector output by the network and the S×S×(B*5+C) dimensional vector corresponding to the real image. The loss function is shown in Equation (2).

[0066]

[0067] Here, coordError, iouError, and classError represent the coordinate error, intersection-over-union (IoU) error, and classification error between the predicted and calibrated data, respectively.

[0068] 2) Sparse map registration of semantic labels is completed through feature matching, and then clustering and segmentation of sparse semantic map points and planar reconstruction are performed. Semantic labels are applied to frequently occurring and relatively stable targets in the scene. In the image, the identified region is represented by Equation (3).

[0069] Region c ={L,{x LT ,h,w}} (3)

[0070] Where c represents x in the image plane LT The coordinates of the top left corner of the semantic region are represented by h and w, which represent the height and width of the semantic region, respectively. L represents the semantic label of this region.

[0071] Accordingly, the three-dimensional semantic map points are represented as Equation (4).

[0072] X = {L, x} (4)

[0073] Where X represents a 3D semantic map point in the world coordinate system, x represents a map depth point, and L represents a semantic label. To obtain X, the feature ID obtained from the feature tracker is used as an anchor point between the image and the sparse depth feature points to determine the 3D semantic point X.

[0074] Semantic annotation projects the semantic points in each image into a 3D space. For a plane in 3D space, the feature points in one image are far from sufficient. Therefore, these semantic points need to be incrementally updated, and the point cloud of these spaces is divided into different planes for planar reconstruction.

[0075] The supervoxel method is used to reduce the number of 3D regions. The supervoxel processing procedure is as follows: an adjacency graph of the voxel cloud is constructed, spatial seeds are generated and filtered, clustering is performed using feature distances, and spatial connectivity between clusters is enhanced through iterative clustering algorithms. This process is repeated on the 3D point cloud at a fixed frequency and with a sliding window to obtain supervoxel clusters for subsequent LCCP segmentation.

[0076] The LCCP method is used to solve the point cloud block segmentation problem based on the concavity-convexity relationship between two adjacent planes. We represent the supervoxels obtained from clustering as Equation (5).

[0077] Region w ={L, N, P center {x0, x1, ..., x} N}} (5)

[0078] Where w represents the world coordinate system, L represents the semantic label of this region, N represents the number of hypervoxels, and P represents the number of hypervoxels. center Let {x0, x1, ..., x} represent the coordinates of the center point of this supervoxel. N} represents all three-dimensional coordinate points within the hypervoxel.

[0079] Subsequently, the Random Sampling Consensus (RANSAC) method is used to compute the regression plane of the 3D semantic supervoxels. The RANSAC process involves randomly selecting 3D semantic feature points, estimating the model, and calculating the number of interior points belonging to this model. This process is repeated iteratively a certain number of times, outputting the plane parameters and the indices of the interior points belonging to the plane. This removes outliers and obtains an accurate semantic plane that satisfies the Region... w The semantic plane is denoted by π. These semantic planes, determined by sparse semantic feature points, constitute the preliminary semantic space as shown in equation (6).

[0080] W′={(L1, π′1), (L2, π′2),…(L n ,π′ n (6)

[0081] Where n is the number of semantic planes.

[0082] 3) A local planar rolling optimization method based on spatiotemporal geometric constraints realizes the reconstruction and perception of semantic scenes: such as Figure 2 As shown, the semantic plane of the input is defined as π. k The number of points falling in the plane is The center coordinates of the plane are The normal vector of the plane is The coefficients of the plane are (A, B, C, D).

[0083] Then, the generated plane is added to the global semantic map. This section divides this step into two sub-steps: insertion and update. Insertion step: In the local plane, using the camera's heading and the normal vectors of other planes, a plane quality loss function with respect to orientation error is established.

[0084]

[0085] in It is the normal vector of all other local planes. It is an index of the local plane. It is the heading vector formed by the coordinates of the camera's head and tail within the sliding window: (like Figure 3 (As shown).

[0086] Update steps: Try to find the plane index closest to the input plane. This process can be represented as shown in equation (8).

[0087]

[0088] Where t is the index number of the plane closest to the newly added plane. If the newly added plane is π... k The number of interior points contained is greater than that of the historical plane π. t Then use π k instead of π t Accordingly, the planar mass loss function for the update step is also defined as equation (9).

[0089]

[0090] in Indicates the proportion of additional interior points. This represents the probability of successfully updating the plane.

[0091] The sum of each error term in the insertion and update is minimized using an unconstrained optimization formula, as shown in Equation (10), to obtain the optimal plane parameters.

[0092]

[0093] The optimization function described above does not contain differential terms, so a direct method is used to solve the function optimization problem. Optimization mainly involves two steps: detection search and pattern shifting. The purpose of detection search is to find a salient point around the initial search point (reference point) to determine a favorable direction. Pattern shifting involves moving the salient point to the new reference point. These two steps are repeated alternately, gradually moving the function to its minimum.

[0094] Finally, all optimized semantic planes were obtained, and the final semantic space was expressed as Equation (11).

[0095] W={(L1, π1), (L2, π2),…(L n , π n (11)

[0096] 4) Construct a carbon semantic map by combining carbon emission monitoring point cloud information. A differential optical depth (DOD) method is used to obtain the spatial concentration of carbon emissions. The quantification index DOD conforms to the following standard, where F is the best-fit transmittance, used to determine the CO2 path density (unit: m^(-2)):

[0097]

[0098] Or convert to path average mixing ratio Unit: (ppm), where r is the target distance measured by the rangefinder, Δσ is the differential absorption cross section of carbon dioxide, and N air It is the air density:

[0099]

[0100] For each DOD value, the model covariance matrix and the residual standard deviation (root mean square error) of the fitted data are proportionally transformed to provide a meaningful confidence estimate. The UAV monitors data at four arbitrary points during flight, acquires the distances to each point, calculates the DOD for each point, and plots the relevant curves. Finally, the concentration of the relevant carbon-containing gas is derived from the DOD and the following formula:

[0101] .

Claims

1. A method for constructing an online carbon semantic map based on sparse fusion, comprising the following steps: 1) Obtain semantic labels for city images based on object detection neural networks; 2) Sparse map registration of semantic labels is completed through feature matching, and then clustering, segmentation and planar reconstruction of sparse semantic map points are performed; 3) A local planar rolling optimization method based on spatiotemporal geometric constraints is used to reconstruct and perceive semantic scenes; 4) Construct a carbon semantic map by combining carbon emission monitoring point cloud information; Step 2) specifically includes: semantically labeling frequently occurring and relatively stable targets in the scene, and representing the identified regions in the image as shown in equation (3). Region c ={L,{x LT ,h,w}} (3) Where c represents x in the image plane LT The coordinates of the top left corner of the semantic region are represented by h and w, which represent the height and width of the semantic region, and L represents the semantic label of this region. Accordingly, the three-dimensional semantic map points are represented by Equation (4): X={L,x} (4) Where X represents a 3D semantic map point in the world coordinate system, x represents a map depth point, and L represents a semantic label; in order to obtain X, the feature ID obtained by the feature tracker is used as the anchor point between the image and the sparse depth feature points to determine the 3D semantic point X. Semantic annotation projects the semantic points in each image into a three-dimensional space. For a plane in a three-dimensional space, the feature points in one image are far from enough. Therefore, it is necessary to incrementally update these semantic points and divide the point cloud of these spaces into different planes according to the plane to perform plane reconstruction. The supervoxel method is used to reduce the number of 3D regions. The supervoxel processing process is as follows: construct the adjacency graph of the voxel cloud, generate and filter spatial seeds, perform clustering using feature distance, and enhance the spatial connectivity between clusters through iterative clustering algorithms; repeat this process on the 3D point cloud at a fixed frequency and in a sliding window manner to obtain supervoxel clusters for subsequent LCCP segmentation. The LCCP method is used to solve the point cloud block segmentation problem based on the concavity-convexity relationship between two adjacent planes. The supervoxels obtained by clustering are represented by Equation (5). Region w (L,N,P center ,{x0,x1,···,x N }} (5) Where w represents the world coordinate system, L represents the semantic label of this region, N represents the number of hypervoxels, and P represents the number of hypervoxels. center This represents the coordinates of the center point of this supervoxel, {x0, x1, ..., x...} N } represents all three-dimensional coordinate points within the supervoxel; Subsequently, the Random Sampling Consensus (RANSAC) method was used to calculate the regression plane of the 3D semantic supervoxels. The RANSAC method process is as follows: randomly select 3D semantic feature points, estimate the model, and calculate the number of interior points belonging to this model; this process is repeated for a certain number of iterations, and the plane parameters and the indices of the interior points belonging to the plane are output. This removes isolated points and obtains an accurate semantic plane that satisfies the Region w The semantic plane is denoted by π. These semantic planes, determined by sparse semantic feature points, constitute the preliminary semantic space as shown in equation (6). W′={(L1,π′1),(L2,π′2),···(L n ,p′ n )} (6) Where n is the number of semantic planes; Step 3) specifically includes: establishing a quality loss function for the semantic plane and proposing a local optimization method for the semantic map based on this loss function. A local plane rolling optimization method based on spatiotemporal geometric constraints realizes the reconstruction and perception of semantic scenes: the input semantic plane is defined as π. k The number of points falling in the plane is The center coordinates of the plane are The normal vector of the plane is The coefficients of the plane are (A, B, C, D); Then, the generated plane is added to the global semantic map; this involves two sub-steps: insertion and update. The insertion step involves establishing a plane quality loss function related to orientation error within the local plane, using the camera's heading and the normal vectors of other planes. in It is the normal vector of all other local planes. It is an index of the local plane. It is the heading vector formed by the coordinates of the camera's head and tail within the sliding window: Update steps: Try to find the plane index closest to the input plane. This process can be represented as shown in equation (8). Where t is the index number closest to the newly added plane; if the newly added plane is π k The number of interior points contained is greater than that of the historical plane π. t Then use π k instead of π t Accordingly, the planar mass loss function for the update step is also defined as equation (9). in Indicates the proportion of additional interior points. This represents the probability of successfully updating the plane. The sum of each error term in the insertion and update is minimized using an unconstrained optimization formula, as shown in Equation (10), to obtain the optimal plane parameters. The above optimization function does not contain differential terms, so the direct method is used to solve the function optimization problem. The optimization mainly includes two steps: detection search and pattern shift. The purpose of detection search is to find a salient point around the initial search point, i.e., the reference point, to determine a favorable direction. Pattern shift is to move the salient point to the new reference point. These two steps are repeated alternately to gradually move the function to the minimum value. Finally, all optimized semantic planes were obtained, and the final semantic space was expressed as Equation (11). W={(L1,π1),(L2,π2),···(L n ,p n )} (11).

2. The online carbon semantic map construction method based on sparse fusion as described in claim 1, characterized in that: Step 1) specifically includes: The object detection network first rescales the image to a uniform resolution input image, then feeds the image into the neural network to predict the location of the target, the confidence of the bounding box containing the target, and the conditional probability of classification. Finally, non-maximum suppression is applied to the results to obtain the classification result with the best probability. The network implementation process is as follows: The input image is divided into S×S grids. Each grid independently predicts B bounding boxes and a confidence score for each bounding box. The confidence score includes the probability that an object exists in the predicted bounding box and the accuracy of the bounding box's location. In addition, each grid also needs to predict a conditional probability C, which is the probability that the object belongs to a certain class given that an object exists in this grid. Therefore, the network output is a vector of S×S×(B*5+C) dimensions. During prediction, the classification information of each grid and the grid confidence score are multiplied to obtain the classification-spatial confidence score of the grid. At the end of the algorithm, a non-maximum suppression method is used to find the bounding box and class of each grid that contains the most accurate target location information. The semantic labels of urban images are obtained based on a target detection neural network. Specifically, the network uses 24 convolutional layers and 2 fully connected layers. It does not require a region proposal network to find targets; target detection and recognition can be completed directly through regression. The convolutional layers are used to extract image features, while the fully connected layers are used to predict the target location and category probability. This network borrows the structure of the GoogLeNet classification network, but unlike GoogLeNet, it uses 1×1 convolutional layers plus 3×3 convolutional layers to replace the Inception module to reduce the dimensionality of computation and normalize the prediction results. The last layer of the network uses a linear activation function, and the linear correction unit ReLU in Equation (1) is used as the activation function for other layers. Furthermore, this network uses mean squared error as the loss function to optimize the model parameters, that is, the mean squared error between the S×S×(B*5+C) dimensional vector output by the network and the S×S×(B*5+C) dimensional vector corresponding to the real image. The loss function is shown in Equation (2). Here, coordError, iouError, and classError represent the coordinate error between the predicted data and the calibration data, the Intersection Over Union (IoU) error, and the classification error, respectively.

3. The online carbon semantic map construction method based on sparse fusion as described in claim 1, characterized in that: Step 4) specifically includes: using a differential optical depth (DOD) method to obtain the concentration of carbon emissions in space; the quantification index DOD conforms to the following standard, where F is the best-fit transmittance, used to determine the CO2 path density, unit: m^(-2): Or convert to path average mixing ratio Unit: ppm, where r is the target distance measured by the rangefinder, Δσ is the differential absorption cross section of carbon dioxide, and N. air It is the air density: For each DOD value, the model covariance matrix and the residual standard deviation (root mean square error) of the fitted data were proportionally transformed to provide a meaningful confidence estimate. The UAV monitored data at four arbitrary points during flight, acquiring the distances between each point, calculating the DOD for each point, and plotting the relevant curves. Finally, the concentration of the relevant carbon-containing gas was derived from the DOD and the following formula:

Images