Urban low-altitude voxelization detection updating method

Abstract

The invention relates to the technical field of urban space element low-altitude detection, and provides an urban low-altitude voxelization detection updating method, which comprises the following steps of: 1, combining urban element voxel perception sensors; step 2, hard synchronization; step 3, outputting a voxel cloud; 4, voxel detection-shielding is carried out; step 5, voxel detection-jitter elimination; step 6, voxel detection-hole filling; step 7, object recognition and extraction; step 8, carrying out element post-processing; and step 9, quality inspection. According to the method, urban low-altitude voxelization detection updating can be well carried out.

Description

Technical Field

[0001] The present invention relates to the technical field of low-altitude detection of urban spatial elements, and specifically, to a method for detecting and updating urban low-altitude voxelization. Background Art

[0002] With the rise of low-altitude economic industries such as drone logistics, urban air traffic, and low-altitude tourism, the demand for high-precision and real-time low-altitude voxel models is increasing day by day to support applications such as flight path planning, obstacle recognition, and airspace management. Urban planners and managers need more accurate and timely low-altitude voxelization information for urban spatial planning, building design review, traffic flow analysis, environmental monitoring, etc. to meet the needs of urban refined management. In emergency situations such as natural disasters and accident disasters, quickly obtaining and updating low-altitude voxelization information is crucial for rescue command, disaster assessment, safety monitoring, etc.

[0003] The main technical methods for rapid and efficient urban voxelization update are the technology of extracting elements based on satellite images and the technology of low-altitude photogrammetry based on drones. The advantages are low cost and high efficiency, but the limitations are also obvious, mainly manifested in: the quality and acquisition frequency of satellite images are affected by weather and cannot be applied to high-frequency application scenarios; the accuracy of extracting urban elements by remote sensing technology based on satellite images depends on the resolution of satellite images, and it faces great challenges in extracting small urban components with high accuracy requirements. The technology of low-altitude photogrammetry based on drones solves the problems of image acquisition frequency and image resolution, but when flying quickly over the photographed area, it faces the problem of misperception and omission of fine information of urban elements caused by ground continuity and the disturbance of moving objects on the ground. Summary of the Invention

[0004] The content of the present invention is to provide a method for detecting and updating urban low-altitude voxelization, which can better detect and update urban low-altitude voxelization.

[0005] According to a method for detecting and updating urban low-altitude voxelization of the present invention, it includes the following steps:

[0006] Step 1: Combine urban element voxel perception sensors, including lidar, inertial navigation IMU, and high-definition cameras, which are fixed on a mounting bracket and mounted on a drone;

[0007] Step 2: Perform hard synchronization, calibrate the parameters of the sensors, establish the spatial attitude conversion relationship between the sensors, and achieve spatial attitude and timing hard synchronization;

[0008] Step 3: Output voxel cloud, receive the perception data of the three sensors, fuse them into a unified spatio-temporal reference, and output a 4D voxel cloud;

[0009] Step 4, Voxel Detection - Occlusion: Establish a voxel occlusion detection model to detect the occluded areas of the elements from consecutive voxel frames and detect and restore the original information of the elements;

[0010] Step 5, Voxel Detection - Jitter Reduction: Establish a voxel jitter detection model to detect the jitter areas of the elements from consecutive voxel frames and detect and restore the original information of the elements;

[0011] Step 6, Voxel Detection - Hole Filling: Establish a voxel missing detection and hole filling model to detect the missing areas of the elements from consecutive voxel frames and detect and restore the original information of the elements;

[0012] Step 7, Object Recognition and Extraction: Use a deep learning model to recognize urban elements, privatize and train the deep learning model based on Yolo for urban elements, and perform three-dimensional urban element object recognition and extraction on the voxel cloud;

[0013] Step 8, Element Post-Processing: Establish an urban element post-processing model, perform different processing flows on different types of elements, and perform normalization processing on urban elements;

[0014] Step 9, Quality Inspection: Conduct quality inspections on the extracted urban elements according to business specifications and quality standards, filter out the elements that do not meet the quality standards, and output a quality report.

[0015] Preferably, in Step 2, taking the IMU coordinate system O i as the reference coordinate system, the lidar coordinate system O l and the camera coordinate system O c are respectively transformed into the IMU coordinate system for spatial reference unification of multi-sensors;

[0016] (2.1) Sensor Calibration;

[0017] The fixed spatial relationship between sensors is calibrated and solved by measuring typical features, and only one solution is required;

[0018] Calibrate through planar features. The equation of the plane π l in the lidar coordinate system O l is:

[0019] ax l +by l +cz l +d = 0

[0020] where: a, b, c are the components of the normal vector, determining the direction of the plane; x l , y l , z l are the coordinates of any point on the plane; d is the constant term, related to the distance from the plane to the origin;

[0021] IMU coordinate system O i corresponding plane π i has the equation:

[0022] Ax i +By i +Cz i +D = 0

[0023] where: A, B, and C are the components of the normal vector, which determine the direction of the plane; x i , y i , z i are the coordinates of any point on the plane; D is the constant term, which is related to the distance from the plane to the origin;

[0024] The extrinsic transformation matrix T li transforms the plane π l to the plane π i , then there is:

[0025]

[0026] where: R is the rotation matrix for the plane transformation, is the translation vector for the plane transformation;

[0027] (2.2) Phase space attitude synchronization;

[0028] At different phases, the conversion relationship between the sensor attitudes O t and the attitude O t+1 remains unchanged. Representing the spatial attitude of the sensor in Euler angles (α, β, γ) and rotating in the order of the Z-axis, Y-axis, and X-axis, the conversion synchronization relationship:

[0029]

[0030] where, R ZYX is the rotation matrix from the phase space O t to the phase space O t+1 , and α, β, r are the Euler angles of the sensor during phase conversion;

[0031] Then, the spatial positioning at the t + 1 phase is derived from the t phase:

[0032] v t+1 = R ZYX (α, β, γ)v t

[0033] where: v t is the rate at the t phase; v t+1 is the rate at the t + 1 phase.

[0034] Preferably, in step 3, the voxel cloud is a 4D data structure, and the multi-dimensional point cloud sliced by time phase is organized in time sequence. The voxel cloud is obtained by fusing the original information sensed by three sensors after the attitude conversion in step 2;

[0035] Voxel cloud organization model:

[0036] C = {c1, c2…c t}

[0037] where c t is the voxel frame at time phase t;

[0038] Voxel organization model:

[0039] c t = {x, y, z, i, r, g, b, o}

[0040] where x is the x coordinate of the voxel, y is the y coordinate of the voxel, z is the z coordinate of the voxel, i is the laser echo intensity of the voxel, r is the R value of the voxel color, g is the G value of the voxel color, b is the B value of the voxel color, and o is the voxel optical flow value.

[0041] Preferably, in step 4, specifically:

[0042] Based on time phase t, extract n consecutive frames of voxels backward from the voxel cloud to obtain the time-sequence voxel dataset C = {C t C t+1 … C t+n}. First, perform plane recognition and segmentation on the t-frame voxels to obtain the plane dataset P = P1 P2 … P n}. Convert the time-sequence voxel set C to the frame plane at time sequence t, project the plane dataset P, calculate the projection distance between the time-sequence frame voxels and the plane, and construct the covariance set F of the voxel to the plane = Perform mutation detection on the covariance dataset. When a mutation is too large, it is determined as occlusion, and the voxels are removed from the time-sequence voxel frames;

[0043] (4.1) Plane equation of the voxel frame:

[0044] ax + by + cz + d = 0

[0045] where {a, b, c} is the normal vector of the plane d is the distance from the plane to the origin;

[0046] Distance from a point to a plane:

[0047]

[0048] where D is the distance from the voxel point to the plane, and {a, b, c} is the plane normal vector;

[0049] (4.2) Voxel frame plane segmentation, calculate the plane normal vector of the nearest neighbor point set of each voxel. When the normal vector changes too much, it is determined as the plane boundary, extract the continuous plane boundary voxels, construct a closed curve, and the voxels inside the curve form a plane;

[0050] p i (x i ,y i ,z i ) and the seed point p s (x s ,y s ,z s )'s normal vector and the included angle θ between:

[0051]

[0052] When the included angle θ is greater than the set threshold, the voxel point is a boundary point.

[0053] Preferably, in step 5, specifically:

[0054] Based on the time phase t, extract the subsequent continuous n frames of voxels from the voxel cloud to obtain the time series voxel data set C = {C t C t+1 … C t+n}, calculate the optical flow for the continuous voxel frames using the method of region feature matching. The voxels with large optical flow are determined as jitter voxels, eliminate the voxels with excessive jitter, and retain the static voxel information in the continuous frames;

[0055] (5.1) Spatiotemporal gradient calculation: For the voxel point (x, y, z, t), the spatiotemporal gradient:

[0056]

[0057] I x (x, y, z, t), I y (x, y, z, t), I z (x, y, z, t), I t (x, y, z, t) respectively represent the spatiotemporal change relationships in the x, y, z, t dimensions;

[0058] (5.2) Based on the spatiotemporal gradient, the optical flow constraint equation for the time phase t is:

[0059] I x (x, y, z, t)u + I y (x, y, z, t)v + I z (x, y, z, t)w = 0

[0060] Among them, u, v, and w are the components of the optical flow in the voxel block method direction, respectively.

[0061] Preferably, in step 6, specifically: after performing plane segmentation on the voxel frames based on step 4, further perform hole detection on the set of segmentation planes P, and find the corresponding voxels in the hole area from the consecutive voxel frames to fill the detected holes to restore the missing voxel information.

[0062] Preferably, in step 7, specifically:

[0063] Object prediction classification uses multi-class label prediction. For each class c i , the predicted probability p i is obtained through logistic regression, and the loss function uses binary cross-entropy loss:

[0064]

[0065] where: Loss class is the object classification prediction loss function; s 2 is the number of classes; is the object classification weight; p i is the actual classification probability; is the predicted classification probability.

[0066] Preferably, in step 8, specifically:

[0067] (8.1) First project the voxel cloud onto a two-dimensional regular grid, and then cluster the irregular and discrete points into a regular voxel grid.

[0068] (8.2) Supervoxel generation: According to the texture, elevation, color, and normal similarity eigenvalue, reassign the voxel points within the grid to generate supervoxels that retain the building boundary features;

[0069] Let the point in the point cloud be p i =(x i , y i , z i ), and the feature vector F i is:

[0070] F i =(T i , E i , C i , N i )

[0071] where T i is the texture feature, E i is the elevation feature, C i is the color feature, N iIt is a normal feature. By calculating the feature distance D between points, point clouds with similar features are clustered into supervoxels:

[0072]

[0073] (8.3) Boundary extraction: Using the least squares method to fit a straight line to the feature distance D, the building boundary is extracted from the voxel cloud, and then constrained and normalized through the spatial geometric features of different urban element types to obtain the final urban element object.

[0074] The beneficial effects of the present invention are as follows:

[0075] By integrating the heterogeneous data of multiple low-altitude sensors and expanding the temporal dimension of spatial data, the present invention can fully detect the spatio-temporal features of urban elements, effectively detect the spatio-temporal correlation of elements, fully retain the spatial and extended attributes of urban elements, improve the recognition and extraction efficiency and accuracy of urban elements, shorten the urban element update cycle and improve the update efficiency. In view of the two major requirements of high precision and high efficiency for urban element update, facing the problems of low cost and frequency update demands, the proposed method and update process of multi-source heterogeneous sensor combination and time-series voxelization detection have the advantages of high element accuracy and low update cost, and have extremely high application value in the field of urban element update technology. Description of the Drawings

[0076] Figure 1 It is a flowchart of a method for urban low-altitude voxelization detection and update in an embodiment. Detailed Embodiments

[0077] To further understand the content of the present invention, the present invention will be described in detail in combination with the drawings and embodiments. It should be understood that the embodiments are only for explaining the present invention and not for limiting it.

[0078] Embodiment

[0079] As Figure 1 shown, the present embodiment provides a method for urban low-altitude voxelization detection and update, which includes the following steps:

[0080] Step 1: Combine urban element voxel perception sensors, including lidar, inertial navigation IMU, and high-definition cameras, which are fixed on a sturdy mount and mounted on a drone.

[0081] Step 2: Hard synchronization. Calibrate the parameters of the sensors, establish the spatial attitude conversion relationship between the sensors, and achieve spatial attitude and time-series hard synchronization.

[0082] In step 2, taking the IMU coordinate system O i as the reference coordinate system, the lidar coordinate system O lConvert to the IMU coordinate system respectively to unify the spatial reference of multi-sensors with the camera coordinate system O; c

[0083] (2.1) Sensor calibration;

[0084] The fixed spatial relationship (external parameters) between sensors is calibrated and solved by measuring typical features, and only one solution is required;

[0085] Calibrate through planar features. The equation of the plane π under the lidar coordinate system O l is: l

[0086] ax l +by l +cz l +d = 0

[0087] where: a, b, c are the components of the normal vector, determining the direction of the plane; x l , y l , z l are the coordinates of any point on the plane; d is the constant term, related to the distance from the plane to the origin;

[0088] The equation of the corresponding plane π under the IMU coordinate system O i is: i

[0089] Ax i +By i +Cz i +D = 0

[0090] where: A, B, C are the components of the normal vector, determining the direction of the plane; x i , y i , z i are the coordinates of any point on the plane; D is the constant term, related to the distance from the plane to the origin;

[0091] The external parameter transformation matrix T li transforms the plane π l to the plane π i , then there is:

[0092]

[0093] where: R is the rotation matrix of the plane transformation, is the translation vector of the plane transformation;

[0094] (2.2) Temporal-spatial attitude synchronization;

[0095] At different times, the sensor attitudes O t and the attitude O t+1 ​​​The conversion relationship remains unchanged. The spatial attitude (α, β, γ) of the sensor is represented by Euler angles. Rotating in the order of the Z-axis, Y-axis, and X-axis, the conversion synchronization relationship is:

[0096]

[0097] Among them, R ZYX is the rotation matrix from the time-phase space O t to the time-phase space O t+1 ; α, β, γ are the Euler angles of the sensor during time-phase conversion;

[0098] Then, the spatial positioning at time-phase t + 1 is deduced from time-phase t:

[0099] v t+1 = R ZYX (α, β, γ)v t

[0100] Among them: v t is the rate at time-phase t; v t+1 is the rate at time-phase t + 1.

[0101] Step 3: Output the voxel cloud, receive the sensing data of three sensors, fuse them into a unified spatio-temporal reference, and output a 4D voxel cloud.

[0102] In Step 3, the voxel cloud is a 4D data structure, and the multi-dimensional point cloud sliced by time-phase is organized in time sequence. The voxel cloud is obtained by fusing the original information sensed by three sensors after the attitude conversion in Step 2;

[0103] Voxel cloud organization model:

[0104] C = {c1, c2…c t}

[0105] Among them, c t is the voxel frame at time-phase t;

[0106] Voxel organization model:

[0107] c t = {x, y, z, i, r, g, b, o}

[0108] Among them, x is the x coordinate of the voxel, y is the y coordinate of the voxel, z is the z coordinate of the voxel, i is the laser echo intensity of the voxel, r is the R value of the color of the voxel, g is the G value of the color of the voxel, b is the B value of the color of the voxel, and o is the optical flow value of the voxel.

[0109] Step 4: Voxel detection - occlusion, establish a voxel occlusion detection model, detect the occlusion area of the elements from the voxel frames of consecutive time-phases, and detect and restore the original information of the elements.

[0110] In Step 4, specifically:

[0111] Based on the time phase t, extract n consecutive frames of voxels from the voxel cloud to obtain the time-series voxel dataset C = {C t C t+1 … C t+n}. First, perform plane recognition and segmentation on the t-frame voxels to obtain the plane dataset P = {P1 P2 … P n}. Convert the time-series voxel set C to the frame plane at time t, project the plane dataset P, calculate the projection distance between the time-series frame voxels and the plane, and construct the covariance set F from the voxels to the plane = Perform mutation detection on the covariance dataset. When a mutation is too large, it is determined as occlusion, and the voxels are removed from the time-series voxel frames;

[0112] (4.1) Plane equation of the voxel frame:

[0113] ax + by + cz + d = 0

[0114] where {a, b, c} is the normal vector of the plane d is the distance from the plane to the origin;

[0115] Distance from a point to a plane:

[0116]

[0117] where D is the distance from the voxel point to the plane, and {a, b, c} is the plane normal vector;

[0118] (4.2) Voxel frame plane segmentation. Calculate the plane normal vector of each voxel's neighbor point set. When the normal vector changes too much (the angle between adjacent normal vectors is too large), it is determined as the plane boundary. Extract the continuous plane boundary voxels and construct a closed curve. The voxels inside the curve form a plane;

[0119] p i (x i ,y i ,z i ) and the normal vector s (x s ,y s ,z s ) of the seed point p and the included angle θ between them:

[0120]

[0121] When the included angle θ is greater than the set threshold, the voxel point is a boundary point.

[0122] Step 5, Voxel Detection - Debounce: Establish a voxel jitter (motion) detection model, detect the jitter (motion) area of the elements from consecutive phase voxel frames, and detect and restore the original information of the elements.

[0123] In Step 5, specifically:

[0124] Taking phase t as the reference, extract n consecutive frames of voxels from the voxel cloud to obtain the time-series voxel dataset C = {C t C t+1 …C t+n}. Calculate the optical flow for the consecutive voxel frames using the method of region feature matching. Voxels with large optical flow are determined as jitter voxels, eliminate the voxels with excessive jitter, and retain the information of the static voxels in the consecutive frames;

[0125] (5.1) Spatiotemporal gradient calculation: For the voxel point (x, y, z, t), the spatiotemporal gradient:

[0126]

[0127] I x (x, y, z, t), I y (x, y, z, t), I z (x, y, z, t), I t (x, y, z, t) represent the spatiotemporal change relationships in the x, y, z, t dimensions respectively;

[0128] (5.2) Based on the spatiotemporal gradient, the optical flow constraint equation at phase t is:

[0129] I x (x, y, z, t)u + I y (x, y, z, t)v + I z (x, y, z, t)w = 0

[0130] where u, v, and w are the components of the optical flow in the normal direction of the voxel block respectively.

[0131] Step 6, Voxel Detection - Hole Filling: Establish a voxel missing detection and hole filling model, detect the missing areas of the elements from consecutive phase voxel frames, and detect and restore the original information of the elements.

[0132] In Step 6, specifically: After performing plane segmentation on the voxel frames based on Step 4, further detect holes in the set of segmented planes P, find the corresponding voxels in the consecutive voxel frames for the detected holes for filling, so as to restore the missing voxel information.

[0133] Step 7, Object Recognition and Extraction: Identify urban elements by a deep learning model, use a deep learning model based on Yolo for private training of urban elements, and perform three-dimensional urban element object recognition and extraction on the voxel cloud.

[0134] In step 7, specifically:

[0135] For object prediction classification, multi-class label prediction is adopted. For each class c i , the predicted probability p i is obtained through logistic regression, and the loss function adopts binary cross-entropy loss:

[0136]

[0137] where: Loss class is the object classification prediction loss function; s 2 is the number of classes; is the object classification weight; p i is the actual classification probability; is the predicted classification probability.

[0138] Step 8, feature post-processing. Establish a post-processing model for urban features, perform different processing procedures on different types of features, and normalize the urban features.

[0139] In step 8, specifically:

[0140] (8.1) First, project the voxel cloud onto a two-dimensional regular grid, and then cluster the irregular and discrete points into a regular voxel grid.

[0141] (8.2) Supervoxel generation: According to the similarity eigenvalue of texture, elevation, color, and normal, reassign the voxel points within the grid to generate supervoxels that retain the building boundary features;

[0142] Let the point p i =(x i , y i , z i ) in the point cloud, and the feature vector is:

[0143] F i =(T i , E i , C i , N i )

[0144] where T i is the texture feature, E i is the elevation feature, C i is the color feature, N i is the normal feature. By calculating the feature distance D between points, cluster the point cloud with similar features into supervoxels:

[0145]

[0146] (8.3)Boundary extraction: Use the least squares method to perform linear fitting on the feature distance D, so as to extract the building boundary from the voxel cloud, and then perform constraint normalization through the spatial geometric features of different urban element types to obtain the final urban element object.

[0147] Step 9, Quality inspection: Conduct quality inspection on the extracted urban elements according to business specifications and quality standards, filter out the elements that do not meet the quality standards, and output a quality report.

[0148] In this embodiment, by integrating the heterogeneous data of multiple low-altitude sensors and expanding the temporal dimension of spatial data, it is possible to fully detect the spatio-temporal characteristics of urban elements, effectively detect the spatio-temporal correlation of elements, fully retain the spatial and extended attributes of urban elements, improve the recognition and extraction efficiency and accuracy of urban elements, shorten the update cycle of urban elements and improve the update efficiency. In response to the two major requirements of high precision and high efficiency for urban element updates, facing the problems of low cost and frequency update demands, the proposed method and update process of multi-source heterogeneous sensor combination and time-series voxelization detection have the advantages of high element accuracy and low update cost, and have extremely high application value in the field of urban element update technology.

[0149] The above schematically describes the present invention and its implementation manners. This description is not restrictive, and what is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the spirit of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for detecting and updating urban low-altitude voxelization, characterized in that: It includes the following steps: Step 1: Combine the urban element voxel perception sensors, including Lidar, Inertial Measurement Unit (IMU), and high-definition camera, fix them on the mounting bracket, and mount them on the unmanned aerial vehicle. Step 2: Perform hard synchronization. Calibrate the parameters of the sensors, establish the spatial attitude conversion relationship between the sensors, and achieve hard synchronization of spatial attitude and time sequence. Step 3: Output the voxel cloud. Receive the perception data of the three sensors, fuse them into a unified spatio-temporal reference, and output the 4D voxel cloud. Step 4: Voxel detection - occlusion. Establish a voxel occlusion detection model, detect the occlusion area of the elements from the continuous temporal voxel frames, and detect and restore the original information of the elements. Step 5: Voxel detection - jitter reduction. Establish a voxel jitter detection model, detect the jitter area of the elements from the continuous temporal voxel frames, and detect and restore the original information of the elements. Step 6: Voxel detection - hole filling. Establish a voxel missing detection and hole filling model, detect the missing area of the elements from the continuous temporal voxel frames, and detect and restore the original information of the elements. Step 7: Object recognition and extraction. Identify urban elements by the deep learning model, use the deep learning model based on Yolo for privatized training of urban elements, and perform 3D urban element object recognition and extraction on the voxel cloud. Step 8: Element post-processing. Establish an urban element post-processing model, perform different processing flows on different types of elements, and perform normalization processing on urban elements. Step 9: Quality inspection. Conduct quality inspection on the extracted urban elements according to business specifications and quality standards, filter out the elements that do not meet the quality standards, and output a quality report.

2. The urban low-altitude voxelization detection and update method according to claim 1, characterized in that: In step 2, with the IMU coordinate system O i as the reference coordinate system, the lidar coordinate system O l and the camera coordinate system O c are respectively transformed into the IMU coordinate system to unify the spatial reference of multiple sensors; (2.1) Sensor calibration; The fixed spatial relationship between the sensors is calibrated and solved by measuring typical features, and only one solution is required. Calibration is performed through planar features, and the lidar coordinate system O l Lower plane π l Equation of: ax l +by l +cz l +d = 0 where: a, b, c are the components of the normal vector, determining the direction of the plane; x l , y l , z l are the coordinates of any point on the plane; d is a constant term, related to the distance of the plane from the origin; IMU coordinate system O i corresponding plane π i Equation: Ax i +By i +Cz i +D = 0 where: A, B, C are the components of the normal vector, determining the direction of the plane; x i , y i , z i are the coordinates of any point on the plane; D is the constant term, related to the distance from the plane to the origin; External parameter transformation matrix T li Transform the plane π l to the plane π i , then we have: Where: R is the rotation matrix for planar transformation, is the translation vector for planar transformation; (2.2) Temporal-spatial attitude synchronization; At different time phases, the sensor attitude O t and the attitude O t+1 have an unchanged conversion relationship. The spatial attitude of the sensor is represented by Euler angles (α, β, γ). When rotating in the order of the Z-axis, Y-axis, and X-axis, the conversion synchronization relationship is as follows: Among them, R ZYX is the phase space O t to the phase space O t+1 rotation matrix, and α, β, r are the Euler angles of the sensor during phase conversion; Then, the spatial positioning at time phase t + 1 is derived from time phase t: v t+1 = R ZYX (α, β, γ)v t where: v t is the rate at time phase t; v t+1 is the rate at time phase t + 1.

3. The urban low-altitude voxelization detection and update method according to claim 2, characterized in that: In Step 3, the voxel cloud is a 4D data structure, and the multi-dimensional point cloud sliced by time phase is organized in time sequence. The voxel cloud is obtained by fusing the original information sensed by the three sensors after the attitude conversion in Step 2. Voxel cloud organization model: C = {c1, c2…c t} Among them, c t is the voxel frame at time phase t; Voxel organization model: c t = {x, y, z, i, r, g, b, o} Among them, x is the x coordinate of the voxel, y is the y coordinate of the voxel, z is the z coordinate of the voxel, i is the laser echo intensity of the voxel, r is the R value of the voxel color, g is the G value of the voxel color, b is the B value of the voxel color, and o is the voxel optical flow value.

4. The urban low-altitude voxelization detection and update method according to claim 3, wherein: In Step 4, specifically: Based on the time phase t, extract n consecutive frames of voxels from the voxel cloud to obtain the time-series voxel dataset C = {C t C t+1 …C t+n}. First, perform plane recognition and segmentation on the t-frame voxels to obtain the plane dataset P = P1 P2…P n}. Convert the time-series voxel set C to the frame plane at time t, project the plane dataset P, calculate the projection distance between the time-series frame voxels and the plane, and construct the covariance set from voxels to the plane Perform mutation detection on the covariance dataset. When a large mutation occurs, it is determined as occlusion, and the voxels are removed from the time-series voxel frames; (4.1) Voxel frame plane equation: ax + by + cz + d = 0 where {a, b, c} is the normal vector of the plane d is the distance from the plane to the origin; Distance from a point to a plane: Among them, D is the distance from the voxel point to the plane, and {a, b, c} is the plane normal vector. (4.2) Voxel frame plane segmentation. Calculate the plane normal vector of the nearest neighbor point set of each voxel. When the change of the normal vector is too large, it is determined as the plane boundary, extract the continuous plane boundary voxels, construct a closed curve, and the voxels inside the curve form a plane. p i (x i ,y i ,z i ) and the seed point p s (x s ,y s ,z s ) of the normal vector and the included angle θ between: When the included angle θ is greater than the set threshold, the voxel point is a boundary point.

5. A method for detecting and updating urban low-altitude voxelization according to claim 4, characterized in that: In Step 5, specifically: Based on the time phase t, extract n consecutive voxels backward from the voxel cloud to obtain the temporal voxel dataset C = {C t C t+1 …C t+n}. Calculate the optical flow for the consecutive voxel frames using the method of region feature matching. Voxels with large optical flow are determined to be jittery voxels, and the voxels with excessive jitter are removed, while the static voxel information in the consecutive frames is retained; (5.1) Spatio-temporal gradient calculation: For the voxel point (x, y, z, t), the spatio-temporal gradient: I x (x, y, z, t), I y (x, y, z, t), I z (x, y, z, t), I t (x, y, z, t) respectively represent the spatio-temporal variation relationships in the x, y, z, t dimensions; (5.2) Based on the spatio-temporal gradient, the optical flow constraint equation at time phase t is: I x (x,y,z,t)u + I y (x,y,z,t)v + I z (x,y,z,t)w = 0 Among them, u, v, and w are the components of the optical flow in the normal direction of the voxel block respectively.

6. The urban low-altitude voxelization detection and update method according to claim 5, characterized in that: In step 6, specifically: after the voxel frames are segmented into planes based on step 4, the segmentation plane set P is further subjected to hole detection, and for the detected holes, the corresponding voxels in the hole areas are found from the consecutive voxel frames and filled to restore the missing voxel information.

7. A method for urban low-altitude voxelization detection and update according to claim 6, characterized in that: In step 7, specifically: Object prediction classification uses multi-class label prediction. For each class c i , the predicted probability p i is obtained through logistic regression, and the loss function uses binary cross-entropy loss: Among them: Loss class is the object classification prediction loss function; s 2 is the number of categories; is the object classification weight; p i is the actual classification probability; is the predicted classification probability.

8. A method for urban low-altitude voxelization detection and update according to claim 7, characterized in that: In step 8, specifically: (8.1) First project the voxel cloud onto a two-dimensional regular grid, and then cluster the irregular and discrete points into a regular voxel grid. (8.2) Supervoxel generation: According to the similarity eigenvalue of texture, elevation, color, and normal, redistribute the voxel points within the grid to generate supervoxels that retain the boundary features of the building. Let the point p in the point cloud i =(x i , y i , z i ), and the eigenvector F i be as follows: F i =(T i ,E i ,C i ,N i ) where T i is the texture feature, E i is the elevation feature, C i is the color feature, N i is the normal feature. By calculating the feature distance D between points, the point clouds with similar features are clustered into supervoxels: (8.3) Boundary extraction: Use the least squares method to perform a linear fit on the feature distance D to extract the building boundary from the voxel cloud, and then perform constraint normalization through the spatial geometric features of different urban element types to obtain the final urban element object.

Images