Highway landform deformation detection method based on unmanned aerial vehicle multi-element perception

Abstract

The present application belongs to the technical field of highway engineering detection and unmanned aerial vehicle remote sensing, and is used to solve the problem that the prior art cannot resist the interference of coordinate drift on spatial correlation due to the dependence on abnormal confidence of post-processing after blocking. Specifically, the present application is a highway landform deformation detection method based on unmanned aerial vehicle multi-element perception, which comprises: obtaining original digital surface models corresponding to a plurality of block data blocks collected by an unmanned aerial vehicle, road skeleton data, and an engineering structure node library; comparing the end mileage of each weak evidence chain candidate with the node mileage in the engineering structure node library; when the mileage start and end interval overlaps or the interval is less than the preset maximum gap, a candidate connection pair is established; comparing the direction symbols of the two candidates in the candidate connection pair, merging to generate a continuous deformation chain segment based on the comparison result, and outputting each continuous deformation chain segment; the present application introduces the engineering structure node library into the cross-block continuity determination, and realizes the clear distinction between physical termination and data flow truncation.

Description

Technical Field

[0001] This invention belongs to the field of highway engineering inspection and UAV remote sensing technology, specifically a highway landform deformation detection method based on UAV multi-sensing. Background Technology

[0002] In UAV-based highway terrain deformation detection, long-distance roads are usually divided into multiple flight segments or tiles for independent data acquisition and processing. Within each segment, digital surface model reconstruction, elevation normalization, and local benchmark fitting are first completed. Then, based on the geometric residuals or anomaly confidence of the boundary overlap area, the detection results of each segment are stitched together to form a deformation distribution map of the entire road. This processing flow implies an assumption: as long as the detection within the segment is reliable enough, continuous deformation across segments can be recovered through boundary adjacency relationships.

[0003] When slow, long-wave subsidence or creep crosses two block boundaries, the independent local benchmark fitting of each block will absorb the same deformation trend as systematic error, resulting in a significant reduction in the elevation residuals on both sides of the boundary. At the same time, the block boundaries often correspond to flight segment switching or battery replacement locations, and their overlap rate and point cloud density are naturally low. The smoothing operation of the boundary buffer further reduces the confidence of the anomaly. In this case, even if the deformation physically crosses the boundary continuously, the anomaly confidence of the block output cannot reflect this continuity. If the stitching process relies solely on the confidence threshold or the reappearance of the anomalous patch to determine the connection, it is very easy to misjudge cross-block deformation as no anomaly or false interruption.

[0004] GPS / inertial navigation drift between adjacent segments may cause the coordinates of the same physical location to be misaligned by several meters. Relying solely on Euclidean distance makes it difficult to distinguish between true continuous deformation and artifacts of different road levels. Furthermore, traditional splicing methods require consistent deformation symbols across blocks, which lacks the capacity to accommodate natural symbol reversals in roadbed wave deformation or bridgehead transition sections, further fragmenting the complete deformation chain.

[0005] Therefore, there is an urgent need for a method that can determine the continuity of cross-block deformation under a block-based independent processing framework. This method should not rely on the abnormal confidence level of post-block processing, should be able to resist the interference of coordinate drift on spatial correlation, and allow the deformation direction to change naturally under certain conditions, thereby avoiding the underreporting and false connections of long waveform deformation. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting highway landform deformation based on UAV multi-sensoring, which solves the problem that existing technologies do not rely on abnormal confidence levels after block post-processing and cannot resist the interference of coordinate drift on spatial correlation. The technical problem to be solved by this invention is: how to provide a highway landform deformation detection method based on UAV multivariate perception that does not rely on the abnormal confidence level of block post-processing and can resist the interference of coordinate drift on spatial correlation.

[0007] The objective of this invention can be achieved through the following technical solutions: A method for detecting highway topographic deformation based on UAV multi-sensor technology includes: Acquire the original digital surface model corresponding to multiple data blocks collected by the UAV, road skeleton data including road centerline mileage and lateral band coding, and engineering structure node library including node mileage and type identifier; The original digital surface model in each data block is projected onto the road skeleton data to generate a high-order sequence sorted by mileage under each band coding. For each high-order segment, calculate the elevation change direction value between adjacent mileages, count the length of segments with the same continuous direction, and mark segments with a length greater than or equal to a preset window as weak evidence chain candidates containing the mileage start and end interval, direction sign, and total elevation change. The termination mileage of each weak evidence chain candidate is compared with the node mileage in the engineering structure node library. When the difference with the first type of node mileage is less than the buffer, it is marked as physical termination and removed. When the difference with the second type of node mileage is less than the buffer, a transition node label is added. For the first and second candidates with the same bit encoding in adjacent blocks and not physically terminated, a candidate connection pair is established when the mileage start and end intervals overlap or the interval is less than the preset maximum gap. Compare the direction signs of the two candidates in the candidate connection pair, merge them based on the comparison results to generate continuous deformation chain segments, and output each continuous deformation chain segment.

[0008] The present invention has the following beneficial effects: 1. When long-wave settlement or creep crosses block boundaries, this scheme does not rely on the abnormal confidence level of post-block processing. Instead, it identifies weak evidence chain candidates by extracting the length of consecutive segments in the same direction from the original high-order sequence. Even if the local benchmark fitting within each block has absorbed the settlement trend as a systematic error and the boundary buffer has reduced the residual to below the traditional threshold, as long as there are a sufficient number of consecutive points of change in the same direction along the mileage direction, the system can still mark them as candidates and enter the cross-block connection process. This reduces the false negative rate of uniform settlement from an uncontrollable state when relying on the confidence threshold to a known range controlled only by the directional stability window length. 2. By introducing the engineering structure node library into the cross-block continuity determination, this scheme achieves a clear distinction between physical termination and data flow truncation. When the ending mileage of a candidate weak evidence chain is close to the bridge expansion joint, tunnel entrance, retaining wall end, or hard rock boundary, the candidate is marked as hard termination and removed from the cross-block connection candidate list, avoiding the generation of false cross-block deformation chains at structural abrupt changes. When the candidate is close to transitional nodes such as cut-fill boundary or soft soil treatment transition section, the node is tagged for subsequent steps to explain the rationality of the direction reversal. This eliminates the ambiguity in traditional methods where physical sudden stops are misjudged as artificial truncation or noise reversals are misjudged as continuous deformation due to fluctuations in boundary confidence. 3. By using skeleton mileage and lateral zone coding instead of absolute geographic coordinates as the benchmark for spatial association, the interference of GPS / inertial navigation drift between flight segments on cross-block matching is decoupled. Whether two candidates establish a connection depends only on the overlap or interval of their mileage intervals in the direction of the road centerline, and whether they belong to the same lateral zone (such as the right shoulder or the left upper slope). Even if adjacent flight segments have a misalignment of several meters in the plane coordinates of the same physical location due to signal loss, as long as their mileages projected onto the skeleton are aligned and their zones are consistent, the system can still connect them correctly. Conversely, if two candidates have overlapping projections on the plane but belong to different zones (such as the main road surface and the slope), no connection is established. This reduces the probability of cross-layer misconnection in coordinate drift scenarios to zero. 4. By introducing a direction reversal tolerance mechanism, this scheme breaks through the hard constraint of deformation symbol consistency in traditional splicing methods. When two candidate direction symbols of adjacent blocks are opposite, the system checks whether there is a transition node label to reasonably explain the reversal, or distinguishes between pseudo reversal and real unreasonable reversal through the consistency check of local boundary direction sequences, and allows a maximum of three reversals in a single deformation chain. As a result, non-rigid continuous deformations such as roadbed wave deformation (alternating settlement and uplift sections) and natural symbol changes of bridge approach slab transition sections can be merged into the same deformation chain output, rather than being fragmented into multiple isolated anomalies. The feedback loop dynamically adjusts the direction stability window length according to the proportion of unexplained reversals, so that the system maintains consistent sensitivity under different data qualities (point cloud density, noise level). Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is the main flowchart of the method in Embodiment 1 of the present invention; Figure 2This is a sub-flowchart of step S2 in Embodiment 1 of the present invention. Detailed Implementation

[0011] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0012] In existing practices of UAV deformation detection for highways, long-distance roads are typically divided into multiple flight segments or tiles for independent processing. Within each segment, elevation normalization and local benchmark fitting are first performed, and then the segments are stitched together based on the residuals or confidence levels of the boundary overlap areas. The underlying logic of this process is that as long as the detection within each segment is reliable enough, continuous deformation across segments can be naturally recovered through boundary adjacency relationships.

[0013] When slow, long-wave subsidence or creep happens to cross the block boundary, the above logic reveals a fundamental contradiction: the local benchmark fitting of each block absorbs the same deformation trend as systematic error, resulting in a significant reduction in the elevation residuals on both sides of the boundary; at the same time, the boundary area already suffers from insufficient point cloud overlap and smoothing filtering operations due to flight segment switching, further weakening the credibility of the anomaly; in this case, even if the deformation physically crosses the boundary continuously, the anomaly confidence level output by the block cannot reflect this continuity. If the stitching process relies solely on thresholds or the reappearance of anomalous patches to determine connectivity, it will inevitably misjudge cross-block deformation as no anomaly or a false interruption.

[0014] More challenging is that GPS / inertial navigation drift between adjacent segments can cause coordinates of the same physical location to be misaligned by several meters. Relying solely on Euclidean distance makes it difficult to distinguish between true continuous deformation and artifacts of different road levels (such as separation of up and down traffic, main road and ramps). Traditional methods require consistent deformation signs across blocks, which lacks the capacity to accommodate natural sign reversals of roadbed wave deformation (settlement in the front section and uplift in the back section) or bridgehead transition sections, further fragmenting the complete deformation chain.

[0015] If the above problems are not resolved, the detection system will continue to lose its ability to distinguish continuous deformation across blocks. Long-wave settlement will be missed due to the lowered confidence level, causing maintenance decisions to miss the optimal window; bridgehead transition sections will be misconnected or disconnected because nodes are located exactly at block boundaries, generating false defect chains; the upper and lower layers of mountain roads will be incorrectly correlated due to coordinate drift, resulting in invalid alarms; ultimately, the efficiency advantage of independent block processing will be completely offset by the missed reports and false alarms in the splicing stage, and the system will be unable to output a continuous deformation distribution map that can be used in the project.

[0016] Example 1: As Figure 1-2As shown, the highway topography deformation detection method based on UAV multi-sensing includes: Step S1: Obtain the original digital surface model corresponding to multiple data blocks collected by the UAV, road skeleton data including road centerline mileage and lateral band coding, and engineering structure node library including node mileage and type identifier; Before executing step S1, this embodiment pre-constructs three types of basic data required for subsequent processing: the original digital surface model, road skeleton data, and engineering structure node library; the construction methods are described below.

[0017] Acquisition of the original digital surface model: The UAV collects lidar point cloud data along the highway corridor in pre-defined flight segments, with each segment corresponding to a data block. For each block, the collected point cloud is filtered for ground points, and a raster-based digital surface model is generated using existing triangular mesh interpolation methods. It is worth emphasizing that this digital surface model has not undergone any detrending or deskewing processing within any block; that is, the original elevation values ​​completely preserve the overall terrain undulations and deformation trends within the block area without any reference surface correction. In this embodiment, the grid spacing is set to 0.2 meters, and each grid point contains planar coordinates and the original elevation value.

[0018] Construction of Road Skeleton Data: Road skeleton data includes the road centerline mileage sequence and lateral band coding rules. The generation method of road centerline mileage is as follows: Obtain the centerline vector data during the highway design phase. This data consists of a series of points sampled at equal intervals, each point carrying planar coordinates. The vector data is then resampled at preset intervals, and the mileage value of each resampled point is calculated using the existing chord length accumulation method. That is, starting from the road starting point, the Euclidean distance between adjacent points is accumulated segment by segment to obtain the mileage value. This forms the road centerline mileage sequence, where each record contains the mapping relationship between the mileage value and the planar coordinates.

[0019] The determination method for lateral band coding is as follows: For any grid point in the original digital surface model, calculate its planar coordinates and the Euclidean distance between it and each point in the road centerline resampling point set. Select the resampling point with the smallest distance as the matching point, and use the mileage value of the matching point as the approximate mileage of the grid point. Calculate the directed lateral offset distance from the grid point to the matching point: Determine the local tangent direction based on the coordinate difference between the two centerline resampling points before and after the matching point. Use the vertical direction of the tangent as the lateral reference, and obtain the offset value using the existing point-to-line directed distance calculation method. The offset value is positive when the grid point is on the right side of the tangent and negative when it is on the left side. Query the preset lateral boundary threshold table based on the offset distance to obtain the band coding of the grid point. This threshold table is defined according to the standard cross-section of the highway. For example, the offset distance for the road surface area is within ±1.5 meters, the left and right shoulders extend outwards by 1.5 to 3.0 meters, and the left and right upper slopes extend outwards by 3.0 to 8.0 meters. The above boundary values ​​can be adjusted according to the actual cross-sectional width for different highway grades.

[0020] Construction of the Engineering Structural Node Library: The engineering structural node library is derived from highway design drawings or as-built survey data. The library is stored in tabular form, with each row recording the mileage location and type identifier of an engineering structural node. Node types are divided into two categories: the first category is hard termination nodes, specifically including bridge abutment expansion joints, tunnel entrances, retaining wall ends, and hard rock interfaces; the second category is transition nodes, specifically including cut-fill interfaces, soft soil treatment transition sections, and pavement structure type change points. The node mileage accuracy requirement is to reach the meter level.

[0021] The execution flow of step S1: After the above three types of data are ready, step S1 officially performs the acquisition operation; reads the original digital surface model file corresponding to each data block from the UAV data storage unit, reads the pre-constructed road skeleton data (including the road centerline mileage sequence and the lateral boundary threshold table) from the geographic information system, and reads the engineering structure node library from the engineering database; the acquired data will be sent to step S2 for subsequent projection and sequence generation processing.

[0022] For ease of understanding, let's take a real highway inspection scenario as an example: A soft soil road section spans from K13+000 to K13+500. The drone collects data in two segments: the first segment covers from K12+800 to K13+200, and the second segment covers from K13+100 to K13+600. The original digital surface model corresponding to the first data block obtained in step S1 contains the original elevation raster within the range of K12+800 to K13+200, and the second block is similar. In the road skeleton data, the centerline mileage value at K13+000 is 13000 meters, and the lateral boundary threshold table uses the aforementioned standard value. The engineering structure node library contains a record: K13+150 is the cut-fill junction (second type node). These data will be used in subsequent steps to detect continuous settlement across the block boundaries.

[0023] Step S2: Project the original digital surface model in each data block onto the road skeleton data to generate a high-order sequence sorted by mileage under each band coding; for each high-order sequence, calculate the elevation change direction value between adjacent mileages, count the length of consecutive segments with the same direction, and mark segments with a length greater than or equal to a preset window as weak evidence chain candidates containing the mileage start and end interval, direction sign and total elevation change. Step S2 follows the three types of data obtained in step S1, and is responsible for mapping the elevation information in the original digital surface model to the road skeleton reference system, and extracting weak evidence chain candidates with directional stability characteristics from the elevation sequence.

[0024] Projection and generation of high-order sequences: For the original digital surface model of each data block, and the road centerline mileage sequence and lateral boundary threshold table constructed in step S1, this embodiment performs the following projection operation.

[0025] For each grid point within the current block, its planar coordinates are denoted as follows: The original elevation value is First, search the resampling points along the road centerline for the point with the closest Euclidean distance to that grid point. ,get Mileage and plane coordinates ;according to Two adjacent resampling points and Determine the local tangent direction Calculate grid points to directional lateral offset distance (Positive values ​​represent the right side of the road, negative values ​​represent the left side); using the lateral boundary threshold table, based on The range of values ​​determines the band code to which the grid point belongs. After completing the above calculations, a record is generated for that grid point. .

[0026] Records of all grid points within the same block are grouped according to band coding; for all records under the same band coding, the mileage values ​​are determined based on a preset distance threshold. (In this embodiment, 0.2 meters) grid points are grouped into a mileage group; the arithmetic mean of the elevation values ​​of all grid points in the mileage group is taken as the representative elevation value at that mileage, and the mileage is rounded to a preset distance threshold. Integer multiples of - that is, taking The mileage points are then sorted in ascending order of their mileage values ​​to form a high-order sequence under this band encoding. The distance between adjacent mileage points in the sequence is strictly equal to If there are no valid grid points within a certain mileage interval, the sequence will have a gap at that position, and no interpolation will be performed to fill it.

[0027] Taking the first block (covering K12+800 to K13+200) in step S1 as an example, assuming that the right shoulder bit code in this block is "R2", the local high-order sequence generated after projection may contain the following points: (12800.0, 45.23), (12800.2, 45.24), (12800.4, 45.25), ..., (13200.0, 46.87); this sequence will be used for directional stability analysis later.

[0028] For gaps appearing in the high-resolution sequence (i.e., the mileage difference between two adjacent valid sampling points is greater than 1.5 times the preset resolution, which is 0.3 meters in this embodiment), the system will forcibly interrupt the directional segment at that location; specifically, if Meters are not counted. And terminate the currently being counted direction segment at Subsequent points from The process begins by re-establishing new paragraphs; this avoids misinterpreting macroscopic fluctuations in discontinuous regions as local deformation directions.

[0029] Calculation of elevation change direction: For each elevation sequence, starting from the second point, calculate the elevation change direction between adjacent mileage points point by point; set a preset minimum threshold. Meters (used to eliminate lidar measurement noise and floating-point operation errors); for the first [unit] in the high-order sequence Points Compared with the previous point Define elevation difference Direction value Assign values ​​according to the following rules: ; +1 indicates an increase, and -1 indicates a decrease. This indicates a balance; for consecutive directional values ​​of 0, they are merged into the previous non-zero directional value; specifically, if the current... And the previous non-zero direction value is Then the direction of the current point is considered If the sequence begins with consecutive directional values ​​of 0, no directional segment is generated until the first non-zero directional value is encountered, at which point the previous 0 values ​​are merged into that directional value; this results in the directional sequence. Each direction value corresponds to the distance from the mileage. arrive The line segment.

[0030] Statistical analysis of stable directional segments and candidate labels for weak evidence chains: left-to-right scan direction sequences Consecutive identical non-zero direction values ​​are merged into a single paragraph; each paragraph records its starting mileage (the mileage of the point preceding the first direction in the paragraph), ending mileage (the mileage of the point following the last direction in the paragraph), direction sign (+1 or -1), and the number of points within the paragraph. (That is, the number of consecutive directional values ​​that make up the paragraph).

[0031] Let the preset window length be (This embodiment initially uses 10 points, with values ​​ranging from a lower limit of 3 to an upper limit of 30, and can be dynamically adjusted in subsequent steps); For each paragraph: like If the condition is not met, the paragraph is marked as a candidate for a weak chain of evidence; at the same time, the total change in elevation of the paragraph is calculated. ,in The elevation of the end point of the paragraph. This is the elevation of the starting point of the paragraph.

[0032] like Then, further check the direction signs of the preceding and following paragraphs (if they exist), as well as the overall direction sign of the short paragraph itself (take the mode of all non-zero direction values ​​within the short paragraph; if all values ​​within the short paragraph are 0, it is considered to be consistent with the direction of the preceding paragraph): If the direction signs of the preceding and following paragraphs are the same, and the direction sign of the short paragraph is the same as that of the preceding and following paragraphs (or all of the short paragraphs are 0), then the points in the short paragraph are considered noise fluctuations and are merged into the preceding and following paragraphs; after merging, the preceding and following paragraphs become a longer paragraph, with its starting mileage taken from the starting mileage of the preceding paragraph and its ending mileage taken from the ending mileage of the following paragraph, and the total elevation change is the sum of the changes in the two paragraphs. If the directional signs of the preceding and following paragraphs are the same, but the directional signs of the short paragraph are opposite to those of the preceding and following paragraphs, then the short paragraph is marked as a direction-flipping transition zone and is not marked as a candidate for a weak chain of evidence. If the directional signs of the preceding and following paragraphs are opposite, then regardless of the direction of the short paragraph itself, the short paragraph will be marked as a direction-flipping transition zone and will not be marked as a candidate for a weak chain of evidence.

[0033] After the above processing is completed, all weak evidence chain candidates are formed into a list, and each candidate contains four fields: starting mileage. End of mileage Direction symbols (Values ​​can be +1 or -1), Total change in elevation ;regardless The absolute value of the size (even if it is only 1 millimeter), as long as it satisfies All were retained.

[0034] Continuing with the previous example, assuming that the elevation difference between adjacent points of the right shoulder within the interval from K12+900 to K12+950 is positive and greater than 0.002 meters, then this section... Reaching the initial window length It was marked as a candidate for a weak chain of evidence, recorded as: starting mileage 12900.0 meters, ending mileage 12950.0 meters, direction +1. Centimeters; If another paragraph has 7 consecutive points with a direction of -1 in the interval K13+100 to K13+120, but the directions of the preceding and following paragraphs are both +1, then the short paragraph is considered as noise fluctuation, and its points are absorbed into the preceding and following rising paragraphs, and no candidate is output separately.

[0035] These weak evidence chain candidates will be sent to step S3 to be compared with the engineering structure node library to determine whether physical termination or additional transition node labels are required.

[0036] Step S3: Compare the end mileage of each weak evidence chain candidate with the node mileage in the engineering structure node library. When the difference with the first type of node mileage is less than the buffer, mark it as a physical termination and remove it. When the difference with the second type of node mileage is less than the buffer, add a transition node label. Step S3 receives the weak evidence chain candidate list output by step S2 and the engineering structure node library constructed in step S1, and determines for each candidate whether it physically terminates at the engineering structure node or whether it needs to be attached with a transition node label for use in subsequent steps.

[0037] Set buffer width In this embodiment, Meters; this value takes into account the positioning error of the engineering structure nodes in the as-built measurement (usually on the order of meters) and the natural transition range of the deformation influence zone; for any weak evidence chain candidate C output in step S2, its starting mileage is denoted as The final mileage is ; Traverse all nodes in the engineering structure node library, and for each node Record its mileage as The type identifier is Determine if the node is located within the candidate mileage interval or the buffer zone: If If the node is related to the candidate, then the node is associated with the candidate; for each associated node, according to The following operations are performed on the category: like For the first type of node (hard termination type, including bridge abutment expansion joints, tunnel entrances, retaining wall ends, and hard rock interfaces), the candidate end mileage is truncated to... The remaining part after truncation (starting mileage) Unchanged, ending mileage is A new candidate is formed, and its total elevation change is calculated proportionally to the length before and after the cutoff. If the remaining length after the cutoff is less than 0.5 meters, the candidate is removed entirely. like For nodes of the second category (transitional type, including cut-fill junctions, soft soil treatment transition sections, and pavement structure type change points), a transitional node label is added to the candidate information, and the mileage of the node is recorded. and type If the candidate interval contains multiple transition nodes, then all of them should be recorded. If no node in the candidate interval meets the conditions, the candidate remains in a state where it can be connected across blocks, and no node information is attached.

[0038] If multiple relevant nodes are detected simultaneously within a candidate interval, they will be processed according to the following priority: First-class nodes (hard termination) take precedence over second-class nodes (transition); the system first selects the first-class node that is closest to the candidate end mileage, and performs truncation or overall removal operation according to the aforementioned rules. The new candidate generated after truncation will no longer contain the truncated segment. Second-class nodes are processed only if there are no first-class nodes (or if there are no first-class nodes in the remaining interval after truncation of the candidate); for all second-class nodes in the remaining interval, all are labeled with transition nodes. If the remaining mileage length of a candidate after hard termination is less than the preset minimum effective length (0.5 meters in this embodiment), the candidate will be removed entirely without any additional labels.

[0039] It should be noted that this step only checks the ending mileage of each candidate. Nearby nodes; candidate starting mileage If a node exists nearby, its impact will be reflected in the processing of the previous adjacent block—because the candidate end mileage of the previous adjacent block will be compared with that node; therefore, nothing will be missed.

[0040] After step S3, each weak evidence chain candidate is given a new status field with three possible values: "hard termination", "transition node" or "no node". Candidates with a status of "hard termination" are isolated and no longer enter the cross-block connection process. Candidates with a status of "transition node" or "no node" continue to step S4.

[0041] Taking the candidate in the previous example (starting mileage 12900.0 meters, ending mileage 12950.0 meters, direction +1) as an example, assuming there are no engineering nodes at K12+950 in the node library, the candidate status is "no node"; the other candidate (with a cut-fill boundary label attached at ending mileage K13+148) has a status of "transition node" and carries a node mileage of 13150 meters; this information will be used in subsequent steps to determine the direction flip tolerance when connecting across blocks.

[0042] Step S4: For the first and second candidates with the same bit encoding in adjacent blocks and not physically terminated, establish a candidate connection pair when the mileage start and end intervals overlap or the interval is less than the preset maximum gap. Step S4 receives the weak evidence chain candidate list processed in step S3. At this time, each candidate has bit encoding, mileage start and end interval, direction symbol, total elevation change and status field ("hard termination", "transition node" or "no node"); the candidate with the status "hard termination" has been removed and does not participate in this step.

[0043] For two adjacent blocks—denoted as block A and block B—to ensure consistency in mileage direction, the system first checks the mileage range of the two blocks: if the minimum mileage of block A is greater than the maximum mileage of block B (i.e., block A as a whole is after block B), then the roles of A and B are swapped to ensure that block A is the block with the smaller mileage (ahead) and block B is the block with the larger mileage (behind) during processing; if the mileage ranges of the two blocks overlap but the order cannot be clearly defined (e.g., interleaved), then all candidate connections between the block pair are skipped to prevent incorrect associations.

[0044] After confirming that the mileage range of block A is entirely before that of block B, the system groups the candidates in block A that are in the state of "transition node" or "no node" according to their band codes, and sorts them in ascending order of starting mileage within the same group; the same processing is performed on the candidates in block B; (subsequent calculations) and The formula remains unchanged, which guarantees that... There may be overlap, so there is no need to consider reverse calculation. In this case, the formula can be simplified: ,when The time interval is positive, otherwise and ).

[0045] Take one candidate from block B Its band encoding is The mileage range is The same applies when searching for the bit-coded portion in the candidate index of block A. Candidate groups; if no such group exists, then Unable to connect to any candidate in block A, skip; Iterate through all candidates in the group. Record all candidate pairs that satisfy the connection condition (i.e. or If there exists at least one condition that satisfies the condition. Then select the best match from them: prioritize the best match. The largest (longest overlapping segment); if there is no overlap ( If so, then select The smallest (closest interval); record the candidate pair corresponding to this best match. And ignore other matches; mileage range is .

[0046] This strategy ensures that when multiple candidates overlap or are closely spaced, the candidate with the strongest spatial correlation is selected for connection, avoiding the artificial truncation of long-distance continuous deformation; if no condition is met... ,but Unable to match.

[0047] Calculate the relationship between two intervals: ; ; in This represents the overlap length (in meters, non-negative) between the two candidates in terms of mileage. This represents the distance between two candidates (in meters, non-negative, positive when there is no overlap, and 0 when there is overlap or they are adjacent); where The maximum gap is preset (10 meters in this embodiment); when and (or When this happens, candidate connection pairs are established and recorded. and The value of each; Only with the first matching condition found Establish connections and use the nearest neighbor priority strategy to avoid ambiguity in multiple matching; unmatched candidates in block A and unmatched candidates in block B are retained as isolated candidates and output as local defects in the final result without cross-block merging.

[0048] Taking the aforementioned soft soil road section as an example, there is a candidate for a weak chain of evidence in the right shoulder of block A. The mileage range is [12900.0, 12950.0]; there is a candidate bit in the same band of block B. Mileage range [12948.0, 13020.0]; Calculate Since the value of 'meter' is greater than 0, a candidate connection pair is established. Meters (indicating an overlap of 2 meters); another example: candidate interval for block A [13100.0, 13150.0], candidate interval for block B [13155.0, 13200.0], then , For distances less than 10 meters, a connection pair should also be established. rice.

[0049] If the candidate interval [13100.0, 13150.0] of block A is different from the candidate interval [13170.0, 13200.0] of block B... If the distance exceeds 10 meters, no connection will be established; if the two candidate bands have different codes, for example, if candidate band A is the right shoulder and candidate band B is the right slope, no connection will be established even if the mileage intervals overlap.

[0050] After step S4, all successfully matched candidate pairs are sent to step S5 for directional gradient tolerance connection and deformation chain merging.

[0051] Step S5: Compare the direction signs of the two candidates in the candidate connection pair, merge them based on the comparison results to generate continuous deformation chain segments, and output each continuous deformation chain segment.

[0052] This step involves several preset parameters, which are defined as follows: Preset maximum number of flips A value of 3 indicates the maximum number of directional flips allowed for a single deformation chain; Preset node association distance The value is 5 meters, used to determine whether the transition node label is close enough to the candidate boundary to explain the direction flip; Preset pseudo-flip detection distance The value is 2 meters, used to determine whether local direction sequence verification is enabled when there are no transition node labels; Preset consistency threshold The value is 0.5 (i.e., 50%), which is used to determine whether the local direction sequence is consistent enough to be considered a false flip; The first threshold for preset window length adjustment The value is 0.2. When the unexplained flip ratio exceeds this value, the window length is increased. The second threshold for preset window length adjustment The value is 0.05. When the unexplained flip ratio is lower than this value, it is included in the window drop count. First step long value The value is taken from 2 sampling points; Second step length value The value is taken as 1 sampling point; Maximum window length The value is taken from 30 sampling points; Window length lower limit The value is taken from 3 sampling points.

[0053] The values ​​of the above preset parameters are based on the following: Preset maximum number of flips. Based on observations and statistics of typical roadbed wave deformation in highway engineering, the directional changes within a single continuous defect section usually do not exceed three times (e.g., settlement-uplift-settlement). If more than three times occur, it is highly likely that different defect events are superimposed. (Preset node association distance) Meters: Considering that the positioning error of structural nodes in as-built measurements is usually on the order of meters, and that there is a transition length of about 2-3 meters on both sides of the deformation-affected zone, 5 meters is sufficient to cover the actual range of the node's influence on deformation; a preset pseudo-flip judgment distance is used. Meters: This value is based on the typical registration error of UAV point clouds in the boundary overlap area (approximately 1-1.5 meters), with a margin added and rounded to 2 meters to ensure that minor misalignments caused by registration deviations will not trigger false flip detection; preset consistency threshold The requirement is that more than half of the positions in the local orientation sequence have the same orientation. This threshold is statistically effective in distinguishing random noise (expected consistency rate of 0.5) from the true trend (consistency rate approaching 1), and allows for a 50% margin of error. The window length adjustment threshold is also required. and Based on engineering experience, when more than 20% of the flips have no reasonable explanation, it indicates that the current window length is too small, causing excessive noise capture; when it is less than 5%, it indicates that the system is in a highly stable state, and the window length can be reduced to improve sensitivity; the first step is to adjust the window length. Second step length value : Employing asymmetric step sizes allows for a more aggressive window raising (to combat noise) than window lowering (to improve sensitivity), thus preventing frequent system oscillations; upper limit of window length. and lower limit The upper limit of 30 corresponds to a continuous unidirectional change of about 6 meters (at a resolution of 0.2 meters), which is sufficient to capture the settlement section of most engineering concerns; the lower limit of 3 corresponds to 0.6 meters, and sections below this value lack statistical significance and are easily polluted by noise.

[0054] Step S5 receives the candidate connection pair list output in step S4, and is responsible for merging the candidate pairs that meet the spatial association conditions into continuous deformation chain segments according to the direction sign relationship, while handling the tolerance judgment of direction reversal and the dynamic feedback adjustment of the preset window length.

[0055] Direct merging when directions are the same: for candidate connection pairs ,remember The direction symbol is (Value can be +1 or -1) The direction symbol is ;like If so, they are directly merged; the starting mileage of the merged deformable chain segment is... Starting mileage The final mileage is End mileage The total change in elevation is ,in and These represent the total elevation changes for the two candidates; if (That is, there is a mileage interval between the two candidates), and this interval area is marked as "discontinuous" in the output results and no elevation interpolation is performed.

[0056] Flip tolerance determination when directions are different: If If the direction flip tolerance determination process is entered, the system will proceed; the system maintains a global variable. Record the number of flips that have occurred in the current deformation chain, initially set to 0, with a maximum allowed value. The default value is 3.

[0057] examine or Has a transition node label been attached (from step S3)? If such a label exists, and the node's mileage location... and End mileage or with Starting mileage The absolute value of the difference is less than the preset node association distance. (In this embodiment, 5 meters is used), then the flipping is determined to be physically plausible; at this point, merging is permitted, and... Increase by 1; if If so, then the connection will be blocked. and Output as two independent deformation chains.

[0058] Example: A candidate Ending at K13+148, and with the added label of a cut-fill boundary transition node (mileage K13+150), another candidate. Starting at K13+152, the direction sign is... Conversely; due to node mileage 150 and The difference between the ending mileage 148 and the current mileage is 2 meters, which is less than 5 meters. This is considered a reasonable rollover, and merging is allowed. The rollover count is incremented by 1.

[0059] False flip determination without transition node labels: If there are no transition node labels, further determine whether to enter the false flip determination process: If the result calculated in step S4 is... (i.e., the two candidates overlap), or (in If the distance is 2 meters (representing the distance for detecting false flips), then the false flip detection process begins; otherwise (i.e., ... and Merging is directly prohibited, and an unexplained flip event is recorded; the specific operation for pseudo-flip detection is as follows: extract The direction sequence of the three valid sampling points at the tail; a valid sampling point refers to a point that actually exists in the high-order sequence of step S2 and belongs to... Paragraph points; from Take the last three points sequentially from the end mileage point towards the start mileage point, and calculate the elevation change direction value between each adjacent pair of points (using the same direction assignment rule as in step S2) to obtain the sequence. ,in correspond The direction of the last line segment; if If the total number of valid sampling points is less than 3, all valid points are taken; if the number of valid points is less than 2 (i.e. only 1 point), no direction value can be calculated. In this case, it is directly determined that local direction consistency verification cannot be performed. This situation is regarded as a false flip condition not being met, and it is treated as an unreasonable flip. Merging and recording unexplained flip events is prohibited.

[0060] Similarly, extract The orientation sequence of the three valid sampling points in the head; from Starting from the initial mileage point and moving towards the end mileage point, take the first three points sequentially, calculate the direction value between each pair of adjacent points, and obtain the sequence. ,in correspond The direction of the first line segment.

[0061] Calculate the number of values ​​with the same direction in two direction sequences; let... for The actual length, for The actual length is taken as ; take the first two sequences Each position is compared one by one, and the number of positions with the same direction is counted. Then the proportion in the same direction ;like If any sequence has no valid direction, it is directly judged as an unreasonable flip. like If there are not enough sample points available for comparison, reliable local orientation consistency verification cannot be performed. Therefore, it is directly judged as an unreasonable flip, and merging and recording of unexplained flip events are prohibited. like Then further judgment: if (Preset consistency threshold) A value of 0.5 (i.e., 50%) is used to determine whether the local direction sequence is sufficiently consistent to be considered a false flip. If so, it is judged as a false flip, and the direction difference is ignored. The merging process does not change the flip count; if... If the flip is deemed genuinely unreasonable, merging is prohibited, and an "unexplained flip event" is recorded. when If there is no transition node label, merging is directly prohibited, and an unexplained flip event is recorded.

[0062] Deformation chain merging and output: All successfully merged candidate pairs are spliced ​​into a complete deformation chain in mileage order; each deformation chain records: mileage range, band code, direction change sequence (e.g., "ascending segment [12900.0, 12950.0] change +3.2cm, descending segment [12952.0, 13000.0] change -1.5cm"), total number of flips, and whether it ends with a hard termination node.

[0063] Feedback adjustment of preset window length: During the execution of step S5, the system counts all directional flip events occurring at each block boundary (including reasonable flips, non-flips after pseudo-flip transformations, and unreasonable flips); two statistics are defined: the number of events at this boundary that are "uninterpreted by transition node labels and judged as unreasonable flips". and the total number of directional flip events on that boundary. (Note: Pseudo-flips are treated as no flips and are not counted.) The system maintains an integer variable. Used to record the number of boundaries that continuously satisfy the window reduction condition, with an initial value of 0; like Then Reset to 0 and skip the window length adjustment step for the current boundary; like If the value is 0, then calculate the ratio. ;when (First threshold) The value is 0.2. When the unexplained flip ratio exceeds this value, the window length is increased. Therefore, the preset window length in step S2 is adjusted. Add 2 sampling points (maximum 30); at the same time Reset to 0; when (Second threshold) The value is 0.05. When the unexplained flip ratio is lower than this value, it will be included in the window drop count. Increase by 1; otherwise ( )Will Reset to 0.

[0064] when When it reaches 5, Reduce the number of sampling points by one, but not below the lower limit of 3, and then... Reset to 0; the above is for The adjustments will take effect in the processing of the next block.

[0065] For scenarios with three or more consecutive blocks, iterative processing is performed in ascending order of mileage. The merged result of the first and second blocks is denoted as chain P2, whose end mileage, end direction, and cumulative number of flips are known. When processing the second and third blocks, P2 is regarded as a virtual candidate from the second block, and it is matched and merged with the candidate with the same bit encoding in the third block according to steps S4 and S5. The flip counter inherits the existing value of P2. If it exceeds the limit, merging is prohibited and P2 is closed as an independent deformation chain. This process is repeated until all blocks are processed.

[0066] This feedback mechanism enables the system to automatically adjust the sensitivity of weak evidence chain extraction based on the noise level and node distribution density in the actual data. When a large number of unreasonable flips appear in the nodeless region, it means that the current window length is too small, and noise fluctuations are misjudged as stable direction segments. Conversely, when almost all flips can be reasonably explained by nodes, appropriately reducing the window length helps to capture weaker continuous deformations.

[0067] This scheme extracts a weak chain of evidence for directional stability by projecting the original digital surface model onto a road skeleton reference frame, replacing the traditional method's reliance on anomaly confidence within blocks. It utilizes an engineering structure node library to determine hard termination and transition node labels for candidates, resolving the challenge of distinguishing between natural termination of physical deformation and artificial truncation of data streams. Through a spatial correlation mechanism between skeleton mileage and lateral band coding, it eliminates the coordinate misalignment caused by GPS drift between flight segments. It allows directional symbols to be flipped within a finite number of times, supplemented by node interpretation, avoiding reliance on wave deformation. Non-rigid defects such as misalignment are addressed; therefore, even if the confidence level is lowered when long-wave settlement crosses block boundaries, this scheme can still control the false negative rate within an acceptable range by extracting continuous directional segments and merging them across blocks; false connections at nodes such as bridgeheads and tunnel entrances are cut off by hard gating, while reasonable flipping of transition zones such as cut-fill junctions is preserved; false connections between upper and lower layers of mountain roads do not occur due to zone isolation; the final output continuous deformation chain segments directly correspond to the actual subgrade response in terms of mileage range, directional sequence, and change amount, providing a quantifiable basis for maintenance decisions.

[0068] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0069] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for detecting highway landform deformation based on multi-sensor recognition by unmanned aerial vehicles (UAVs), characterized in that, include: Acquire the original digital surface model corresponding to multiple data blocks collected by the UAV, road skeleton data including road centerline mileage and lateral band coding, and engineering structure node library including node mileage and type identifier; The original digital surface model in each data block is projected onto the road skeleton data to generate a high-order sequence sorted by mileage under each bit code. For each high-order segment, calculate the elevation change direction value between adjacent mileages, count the length of segments with the same continuous direction, and mark segments with a length greater than or equal to a preset window as weak evidence chain candidates containing the mileage start and end interval, direction sign, and total elevation change. The termination mileage of each weak evidence chain candidate is compared with the node mileage in the engineering structure node library. When the difference with the first type of node mileage is less than the buffer, it is marked as physical termination and removed. When the difference with the second type of node mileage is less than the buffer, a transition node label is added. For the first and second candidates with the same bit encoding in adjacent blocks and not physically terminated, a candidate connection pair is established when the mileage start and end intervals overlap or the interval is less than the preset maximum gap. Compare the direction signs of the two candidates in the candidate connection pair, merge them based on the comparison results to generate continuous deformation chain segments, and output each continuous deformation chain segment.

2. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The original digital surface model is the original reconstructed elevation data without intra-block detrending processing; the road centerline mileage in the road skeleton data is generated in the following way: the road centerline vector data is resampled at a preset sampling interval, and the mileage value of each resampled point is calculated by accumulating the chord length between adjacent points; the lateral banding code is determined in the following way: the directional lateral offset distance of the grid point relative to the nearest centerline point is calculated, and the corresponding banding code is obtained by querying the preset lateral boundary threshold table based on the offset distance. The lateral boundary threshold table defines the offset distance intervals corresponding to the road surface, shoulder and slope.

3. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The steps for generating a high-order sequence sorted by mileage under each band encoding include: for grid points with the same band encoding within the same data block, grouping multiple grid points with mileage differences within a preset distance threshold into a mileage group, calculating the arithmetic mean of the elevation values ​​of all grid points in the mileage group as the representative elevation value at that mileage, and rounding the mileage to an integer multiple of a preset resolution; sorting the rounded mileage points in ascending order of their mileage values ​​to obtain the high-order sequence, wherein the spacing between adjacent mileage points in the sequence is the preset resolution.

4. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The steps for calculating the elevation change direction value between adjacent mileages specifically include: setting a preset minimum threshold; for the elevation difference between the i-th point and the (i-1)-th point in the elevation sequence, when the elevation difference is greater than the preset minimum threshold, the direction value is assigned as the first direction value indicating an increase; when the elevation difference is less than the negative preset minimum threshold, the direction value is assigned as the second direction value indicating a decrease; when the absolute value of the elevation difference is less than or equal to the preset minimum threshold, the direction value is assigned as the third direction value indicating a flat position; merging consecutively occurring third direction values ​​into the preceding non-zero direction value; if consecutive third direction values ​​appear at the beginning of the sequence, no direction segment is generated until a non-zero direction value is encountered.

5. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The step of counting the lengths of consecutive paragraphs in the same direction and marking paragraphs with a length greater than or equal to a preset window as weak evidence chain candidates further includes: for short paragraphs with a length less than the preset window, if the direction signs of the preceding and following paragraphs are the same and the mode of the direction values ​​in the short paragraph is the same as the direction signs of the preceding and following paragraphs, then the points in the short paragraph are merged into the preceding and following paragraphs; if the direction signs of the preceding and following paragraphs are the same but the mode of the direction values ​​in the short paragraph is opposite to the direction signs of the preceding and following paragraphs, or the direction signs of the preceding and following paragraphs are opposite, then the short paragraph is marked as a direction-flipping transition zone and is not marked as a weak evidence chain candidate; the total elevation change of each weak evidence chain candidate is equal to the elevation value of the end point of the paragraph minus the elevation value of the start point of the paragraph.

6. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The node type identifiers in the engineering structure node library include a first type of node and a second type of node. The first type of node is a hard termination node, specifically including bridge abutment expansion joints, tunnel entrances, retaining wall ends, and hard rock interfaces. The second type of node is a transition node, specifically including cut-fill interfaces, soft soil treatment transition sections, and pavement structure type change points.

7. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The steps for merging and generating continuous deformation chain segments based on comparison results specifically include: when the direction symbols of the first candidate and the second candidate in the candidate connection pair are the same, they are directly merged. The mileage start and end interval of the merged pair extends from the starting mileage of the first candidate to the ending mileage of the second candidate, and the total elevation change is the sum of the total elevation change of the first candidate and the total elevation change of the second candidate; when the direction symbols of the two candidates are different, it is checked whether the first candidate or the second candidate has an attached transition node label. If so, and the difference between the node mileage and the candidate's ending mileage or starting mileage is less than the preset node association distance, then merging is allowed and the cumulative number of flips is incremented by one; if there is no transition node label and the mileage intervals of the two candidates overlap or the interval distance is less than or equal to the preset pseudo-flip judgment distance, the direction sequence of a preset number of sampling points at the tail of the first candidate and the direction sequence of a preset number of sampling points at the head of the second candidate are extracted. The proportion of the same direction value in the two direction sequences is calculated. When the proportion exceeds the preset consistency threshold, it is considered a pseudo-flip and merged in the same direction manner. Otherwise, merging is prohibited and an unexplained flip event is recorded.

8. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 7, characterized in that, The preset maximum value of the cumulative number of flips is the preset upper limit of the number of flips; the specific method for extracting the direction sequence of the preset number of sampling points of the first candidate tail is as follows: take the last preset number of valid sampling points in sequence from the end mileage point of the first candidate to the start mileage direction, and calculate the elevation change direction value between each adjacent point respectively; the specific method for extracting the direction sequence of the preset number of sampling points of the second candidate head is as follows: take the initial preset number of valid sampling points in sequence from the start mileage point of the second candidate to the end mileage direction, and calculate the elevation change direction value between each adjacent point respectively.

9. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 7, characterized in that, It also includes the step of dynamically adjusting the length of the preset window: calculating the ratio of the number of orientation flip events without transition node label interpretation on the boundary of each block to the total number of orientation flip events; When the ratio is greater than the first preset threshold, the length of the preset window is increased by the first step length value; when the ratio is less than the second preset threshold and the ratios of a consecutive preset number of block boundaries are all less than the second preset threshold, the length of the preset window is decreased by the second step length value; the length of the preset window is maintained between the lower limit value and the upper limit value.

10. The method for detecting highway landform deformation based on UAV multi-sensing according to claim 1, characterized in that, The step of projecting the original digital surface model in each data block onto the road skeleton data further includes: for each grid point in the original digital surface model, calculating the Euclidean distance between its planar coordinates and each point in the road centerline resampling point set, selecting the resampling point with the smallest distance as the matching point, using the mileage value of the matching point as the mileage value of the grid point, and using the directional distance of the line connecting the grid point to the matching point as the lateral offset distance, wherein the direction is determined by the coordinate difference between the two centerline resampling points before and after the matching point.

Images