Intelligent inspection trolley for overhead contact system 1c and 4c

Abstract

The present application belongs to the technical field of contact network inspection, and particularly relates to a contact network 1C and 4C comprehensive intelligent inspection trolley, which comprises a trolley body, a data acquisition device and a data processing device; wherein the data acquisition device synchronously acquires point observation information and image observation information along a contact network corridor during the trolley body advancing, so as to jointly constitute a unified corridor reference map with a track center line, a track surface reference plane, a wire reference band and a support calibration point column; the data processing device adopts a domain Hough transform guided by the corridor reference map and a relationship backfill algorithm to identify and locate a contact network positioner, a dropper and a sectional insulator under the constraint of the corridor reference map, and finally generates a positioning landing point and a stable pose of a corresponding target in the corridor reference map. The present application effectively improves the detection precision, stability and processing efficiency, reduces the requirement for manual intervention, and significantly improves the maintenance quality and operation efficiency of the railway contact network.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of contact network inspection, and particularly relates to a 1C and 4C comprehensive intelligent inspection trolley for a contact network. BACKGROUND

[0002] The contact network is an important component of the traction power supply system of an electrified railway, mainly composed of wires, droppers, positioners, support columns, insulators and other accessories, and is continuously erected above the train operation along the track. The state of the contact network directly affects the safety and stability of train operation, so it needs to be periodically inspected. The traditional inspection method mainly relies on manual inspection and handheld measuring equipment, and the maintenance personnel need to enter the line area during the train operation gap to observe and measure. This method is limited by environmental conditions, maintenance personnel's experience and safety factors, and has problems such as low efficiency, high labor intensity, poor data consistency, and cannot realize large-scale and rapid inspection under high-density traffic conditions.

[0003] To solve the problem of insufficient efficiency of manual inspection, some existing technologies introduce video acquisition devices and laser ranging devices installed on track cars or engineering vehicles, which collect images and spatial data during the vehicle running process, and then manually or semi-automatically analyze the data. For example, some systems use industrial cameras to obtain panoramic images of the contact network, and extract the outlines of wires, positioners and other components through image processing methods; some systems use two-dimensional laser scanners to obtain the relative position data of support columns and wires, which are used to judge the geometric state. However, these systems usually use a single sensor or independent data processing flow, and do not fully integrate different types of observation data, which leads to false positives or missed detections in complex backgrounds. At the same time, for local structures such as positioners, droppers and sectional insulators, the detection effect is significantly reduced when the image features are not obvious, the lighting conditions are poor or they are blocked by other components. SUMMARY

[0004] Therefore, the main purpose of the present application is to provide a 1C and 4C comprehensive intelligent inspection trolley for a contact network, which fuses the point-like observation information of a three-dimensional laser radar and the image observation information of an industrial camera, establishes a track center line, a track surface reference plane, a wire reference band and a support column calibration point column and other spatial priors, and realizes accurate detection of the contact network and its attached components. The system uses a method combining domain Hough mapping and spatial constraints to accumulate and filter features in specific domains such as wires, arc-shaped elements and cross elements, and through relationship backfilling and peak decoupling mechanisms, ensures that key components such as positioners, droppers and sectional insulators are stably identified and positioned in long-distance inspection. This technical solution effectively improves the detection accuracy, stability and processing efficiency, reduces the need for manual intervention, and significantly improves the maintenance quality and operation efficiency of the railway contact network.

[0005] The technical scheme adopted by the present application is as follows:

[0006] The contact net 1C and 4C comprehensive intelligent inspection trolley comprises a trolley body, a data acquisition device and a data processing device; wherein the data acquisition device synchronously obtains point observation information and image observation information along the contact net corridor during the trolley body advancing, and an intensity projection graph obtained by projecting the point observation information and pixel-to-pixel aligned with the image observation information, determines a track surface reference plane according to the point observation information, determines a wire reference band according to the image observation information, and determines a column calibration point column according to the point observation information and the image observation information, so as to jointly constitute a unified corridor reference graph with the track center line, the track surface reference plane, the wire reference band and the column calibration point column; the data processing device adopts a domain Hough transform guided by the corridor reference graph and a relationship backfill algorithm, identifies and locates the contact net positioner, the dropper and the sectional insulator under the constraint of the corridor reference graph, finally generates the positioning landing point and the stable pose of the corresponding target in the corridor reference graph, and calculates the lateral offset value, the wire longitudinal position and the hanging point section number according to the positioning landing point; the height value, the lateral offset value, the wire longitudinal position, the hanging point section number and the stable pose are checked one by one for category consistency, position consistency and connection consistency, if consistent, the geometric parameter result set is written, and if inconsistent, the abnormal record set is registered.

[0007] Further, the data processing device projects each positioning landing point in the track surface reference plane to the track center line to obtain a mileage positioning point, takes the reading of the trolley mileage meter at the mileage positioning point as the longitudinal coordinate, and records the shortest horizontal distance projected to the track center line as the lateral offset value, projects each positioning landing point to the wire reference axis along the direction perpendicular to the wire reference axis in the track surface reference plane to obtain a wire correlation point; reads the interval of the positioning landing point in the vertical direction relative to the wire correlation point as the height value, reads the sequential position of the positioning landing point along the wire reference axis as the wire longitudinal position, and finds the nearest two calibration points in the column calibration point column, positions the mileage positioning point between the two calibration points, and determines the hanging point section number according to the distance proportion to the two points.

[0008] Further, the point-like observation information collected by the data collection device is a discrete three-dimensional point set output by the three-dimensional laser radar during the movement of the trolley, each point containing spatial position and reflection intensity; the image observation information is a sequence of gray-scale images obtained by an industrial camera mounted on the trolley in the moving direction; two rail top lines are extracted from the point-like observation information, and the midlines of the two rail top lines are taken as the track center lines; a plane determined by the upper surfaces of the two rails is taken as the track reference plane; in the high-position slice of the point-like observation information, a preliminary guide wire trajectory extending along the track is obtained according to the continuity of linear features, and a guide wire reference band is formed above and below the preliminary guide wire trajectory with the preliminary guide wire trajectory as the center line; and a columnar point cluster or a stripe extending along the vertical direction and having an approximately constant cross section is detected in the point-like observation information and the intensity projection map, and a projection sequence formed by the columnar point cluster or the stripe on the track center line is recorded as a support post calibration point column.

[0009] Further, the data processing device comprises a candidate edge construction unit, a domain Hough mapping unit, a guide wire reference axis establishment unit, a relationship backfilling unit, a peak decoupling output unit, and a result output unit; wherein the candidate edge construction unit takes the point-like observation information and the image observation information as input, performs edge extraction and projection alignment to form a candidate edge set for subsequent voting; the domain Hough mapping unit performs Hough mapping on the candidate edge set in at least one Hough domain to generate at least one cumulative graph; wherein the Hough domain includes any one or more of a linear domain, an arc-shaped domain, or a crossing domain; the whole Hough mapping process is spatially constrained by the guide wire reference band and the support post calibration point column, and only candidate edges compatible with the two are allowed to produce valid votes; the guide wire reference axis establishment unit searches for a stable peak column in the cumulative graph along the moving direction and connects it as a guide wire reference axis, which corresponds one-to-one to the guide wire reference band and serves as an attachment object for subsequent relationship judgment and positioning; the relationship backfilling unit writes the connection relationship between the candidate and the guide wire reference axis and the support post calibration point column into the cumulative graph according to a template dictionary containing a catenary positioner, a dropper, and a sectional insulator, performs relationship backfilling and accompanying component constraint, and obtains a candidate set of three types of targets; the peak decoupling output unit performs peak decoupling processing on peaks coupled due to adjacent components, and performs cross-slice aggregation on continuous slices according to category consistency, direction stability, and connection sequence stability to form a robust peak column, and outputs the positioning landing point and the stable pose of each target.

[0010] Further, the specific process performed by the candidate edge construction unit includes: performing edge detection on each frame of gray-scale image in the image observation information and the corresponding intensity projection map to obtain an edge pixel set; slicing the point-like observation information at a height interval and projecting it onto the corridor reference plane, and extracting the jump boundary using a neighborhood difference method to obtain a slice boundary point set; and combining the edge pixel set and the slice boundary point set into a candidate edge set.

[0011] Further, the execution process of the domain-specific Hough mapping unit includes: performing line domain Hough mapping on the candidate edge set in the corridor reference map, establishing a line domain accumulation map containing only angle scale and distance scale, and using the conductor reference band as a mask to allow only the candidate edges consistent or tangent to the conductor reference band to vote; performing arc domain Hough mapping on the candidate edge set in the same region, establishing an arc domain accumulation map containing angle scale and curvature level, for responding to short elements in the insulator profile in the form of arc or bending; performing cross domain Hough mapping on the candidate edge set, establishing a cross domain accumulation map recording the intersection relationship of two line segments, for detecting short line elements intersecting or tangent to the conductor.

[0012] Further, the execution process of the relationship backfilling unit includes: in the narrow band on both sides of the conductor reference axis, finding the short line peaks intersecting or tangent to the conductor reference axis by using the cross domain accumulation map, and backfilling the votes to the line domain accumulation map; when the intersection points of the short line peaks present a stable order on the conductor reference axis, confirming the corresponding locator candidate; taking each candidate connection point on the conductor reference axis as the starting point, searching for a near-vertical peak column in the line domain accumulation map along the normal direction in the rail surface reference plane, and verifying its continuity in adjacent slices; the verified peak column is taken as a candidate for the pendant; segmented insulator relationship backfilling: in the sliding window along the conductor reference axis, the line domain accumulation map and the arc domain accumulation map are combined to find a sequence group composed of multiple short line peaks and an arc-shaped response; when the sequence group meets the conditions of being located on the conductor reference axis, the short line peaks and the arc-shaped response appearing alternately, and the connection order being consistent with the template dictionary, it is confirmed as a segmented insulator candidate; the locator candidate, the pendant candidate and the segmented insulator candidate are combined to form a candidate set of three types of targets.

[0013] Further, the execution process of the peak decoupling output unit includes: for the peak coupling caused by adjacent components in the same region, performing sub-direction cluster splitting: dividing the source edges into multiple sub-clusters according to the angle scale in the peak neighborhood, independently reconstructing their sources respectively, and selecting the main peak according to the connection consistency with the conductor reference axis, and retaining the rest as secondary peaks; on the continuous slices along the line, the candidates of each type are summarized according to the category consistency, direction stability and connection order stability to form a robust peak column; performing micro-Hough iteration on the robust peak column in the original observation, gradually narrowing the support edge set and refining the axis or contour, and finally generating the positioning landing point and stable pose of the corresponding target in the corridor reference map; the original observation is a joint data set formed by aligning the point-like observation information, image observation information and intensity projection map at the same advancing moment.

[0014] By adopting the above technical solutions, the following beneficial effects are achieved: high-precision and full-automatic detection of the overhead line system and its accessory components during train running is achieved. By establishing a track center line and a track surface reference plane, and using a wire reference band and a support calibration point column as spatial prior constraints, the application can effectively reduce the search space and reduce irrelevant background interference. In the candidate edge construction process, the fusion extraction of the point cloud slice boundary and the image edge makes the feature information more complete, significantly improving the robustness of the detection. The domain-specific Hough mapping performs cumulative operation in three specific domains of the wire, arc-shaped component and cross component, and filters invalid votes in combination with the spatial constraint condition, so that the target response can still be quickly and accurately obtained in a complex scene. The introduction of the relationship backfill unit enables the spatial and structural relationship between the positioner, dropper and segmented insulator and other components to be dynamically completed and verified, avoiding missed detection and false detection caused by insufficient single-frame features. The peak decoupling and cross-slice aggregation mechanism can eliminate the interference between the responses of adjacent components and maintain the stability and continuity of the detection results along the line direction. The overall technical scheme of the application not only improves the precision and stability of the detection, but also greatly outperforms the traditional single-domain detection and post-processing methods in terms of processing speed, and can adapt to long-distance and high-frequency inspection tasks, helping to change the inefficient mode of railway overhead line system maintenance from relying on manual work to intelligent and automated high-reliability mode. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The data acquisition device and data processing device of the overhead line system 1C and 4C comprehensive intelligent inspection trolley provided by the embodiment of the application are shown in the component structure diagram.

[0016] Figure 2 The Hough transform cumulative graph peak response curve provided by the embodiment of the application is shown in the schematic diagram.

[0017] Figure 3 The peak decoupling before and after the signal comparison experiment curve provided by the embodiment of the application is shown in the schematic diagram. DETAILED DESCRIPTION

[0018] All features disclosed in this specification, and / or the steps of any method or process specified in this specification, can be combined in any combination, except combinations where at least some of the features and / or steps are mutually exclusive.

[0019] Any feature disclosed in this specification, unless stated otherwise, can be replaced by alternative features or equivalents having the same or a similar effect. That is, unless stated otherwise, each feature is one example only of a number of alternative or comparable features.

[0020] REFERENCE Figure 1 The overhead line system 1C and 4C comprehensive intelligent inspection trolley comprises a trolley body, a data acquisition device and a data processing device.

[0021] The trolley body in this embodiment is used to carry the data acquisition device and the data processing device and to travel stably on the railway line, so as to meet the requirements of the contact net 1C and 4C comprehensive intelligent inspection trolley in synchronously acquiring point observation information and image observation information and generating a strength projection graph during the travel. The trolley body includes a load-bearing structure composed of a whole welded frame and replaceable end beams at both ends. A load-bearing wheel set and a guide wheel set matched with the track gauge are arranged below the load-bearing structure. The load-bearing wheel set bears the weight and the traction force transmission, and the guide wheel set limits the lateral deviation and cooperates with the anti-falling mechanism to prevent derailment. The load-bearing wheel shaft end is equipped with a trolley odometer. The output of the trolley odometer is used as a reference input of the longitudinal position and speed for the data processing device. The upper surface of the load-bearing structure is arranged with a vibration damping mounting seat for the three-dimensional laser radar and the industrial camera. The vibration damping mounting seat adopts a layered vibration isolation and limiting combined structure, which ensures that the scanning window of the three-dimensional laser radar and the imaging field of view of the industrial camera remain stable under the travel condition and do not block the space region where the wire reference band and the support calibration point column are located. In order to ensure the line-of-sight projection relationship between the point observation information and the image observation information, the trolley body is provided with a rigid relative positioning frame between the three-dimensional laser radar and the industrial camera. A mechanical reference surface and an alignment pin are arranged on the frame, which are used to fix the relative attitude of the two types of sensors and to reposition after maintenance and replacement.

[0022] A rainproof fairing and a transparent protective window are arranged at the front end of the trolley body. The transparent protective window is made of anti-scratch material and is equipped with an automatic cleaning assembly and a defogging heating assembly, which ensures that the image observation information and the strength projection graph are not affected by pollution under night and bad weather conditions. The inside of the frame is arranged with a power supply system, a heat dissipation air duct and a cable channel. The power supply system provides continuous and stable power supply for the three-dimensional laser radar, the industrial camera and the data processing device. The heat dissipation air duct guides the cooling air flow to the heat concentration area of the data processing device and the power supply system through the guide plate. The trolley body shell is provided with multiple levels of sealing to improve the dustproof and waterproof capability. In order to obtain the required visible range of the track reference plane and the track center line, the trolley body is provided with an observation window outside the load-bearing wheel set, so that the three-dimensional laser radar can collect two groups of rail surface points in the low elevation direction, and the industrial camera can collect the texture related to the rail edge in the gray image sequence; on this basis, the data acquisition device divides the two groups of rail surface points according to the point observation information and determines the track reference plane and the track center line, so as to ensure the reference consistency of the corridor reference graph. The trolley body is provided with a three-dimensional laser radar high mast and an industrial camera transverse beam at the top. The three-dimensional laser radar high mast is used to expand the coverage of the point observation information in the height range of the wire reference band and the sectional insulator. The industrial camera transverse beam is used to adjust the field of view without changing the imaging internal parameter, so that the image observation information stably covers the wire reference band and the support calibration point column.

[0023] In order to improve the quality of night imaging, visible light illumination components and narrowband light supplement components are arranged on both sides of the trolley body, and synchronization is realized through the trigger line with the industrial camera, so that the image observation information is kept in time sequence when the light supplement is turned on and off, facilitating the one-to-one alignment of the intensity projection graph and the image observation information. The trolley body is provided with a time reference distribution line and a trigger distribution line, which simultaneously distribute the unified time reference and trigger signal to the three-dimensional laser radar, industrial camera and trolley odometer collection unit, so that the point observation information, image observation information and intensity projection graph are aligned at the same advancing time, ensuring that the construction of the corridor reference graph, the generation of the candidate edge set, the establishment of the guide reference axis and the subsequent calculation of the positioning drop point and stable pose are all performed with consistent time and space reference. In order to reduce the influence of running vibration on point observation information and image observation information, a suspension buffer component is arranged between the load wheel group and the frame of the trolley body, and a two-stage vibration isolation structure is arranged at the installation position of the three-dimensional laser radar and the industrial camera. The suspension buffer component mainly absorbs track joint impact and turnout impact, and the two-stage vibration isolation structure suppresses high-frequency micro-vibration, so that the edge picking process of the candidate edge set remains stable. A maintenance cabin and a data interface panel are arranged at the tail end of the trolley body. The maintenance cabin is used for daily cleaning and replacement of the three-dimensional laser radar transparent protective window, industrial camera lens protection cover and illumination component. The data interface panel provides functions such as power on, power off, time setting, trigger, data export and self-checking, facilitating the completion of coordinate alignment check of the corridor reference graph before leaving the warehouse. The trolley body is provided with installation reference hole series and calibration mark stickers in the structural design, so that after replacing the three-dimensional laser radar or industrial camera, the reference hole and the mark sticker can be used to complete fast geometric alignment, ensuring that the one-to-one alignment relationship of the intensity projection graph and the image observation information is restored to the factory state.

[0024] The data acquisition device is installed on the trolley body of the contact net 1C and 4C comprehensive intelligent inspection trolley, which synchronously acquires point observation information and image observation information at the same advancing time under the unified time reference and trigger control, and generates an intensity projection graph according to the line-of-sight projection relationship between the two, thereby providing aligned data basis for the construction of the corridor reference graph and subsequent identification and positioning. The point observation information is a discrete three-dimensional point set output by the three-dimensional laser radar during the advancing process of the trolley body, each point containing spatial position and reflection intensity; the image observation information is a gray image sequence acquired by the industrial camera in the advancing direction; and the intensity projection graph is obtained by projecting the point observation information onto the imaging plane of the industrial camera according to the line-of-sight and aligning it with the image observation information pixel by pixel.

[0025] In the implementation process, the data acquisition device first establishes the fixed relative pose between the three-dimensional laser radar and the industrial camera through rigid installation and geometric alignment, and through the time reference distribution and trigger distribution of the trolley body, the three-dimensional laser radar scanning period, the industrial camera exposure time and the trolley odometer mileage record form one-to-one corresponding time stamps, so as to ensure that the point observation information, the image observation information and the intensity projection graph are aligned within the same range of travel mileage. In order to ensure that the image observation information still has available contrast at night and in bad weather, the data acquisition device and the visible light illumination component and the narrowband light supplement component of the trolley body work synchronously, and through the cooperation of the trigger line and the exposure time of the industrial camera, the light supplement state is written into the time stamp of the image observation information. After receiving the point observation information, the data acquisition device first re-collects the low-elevation-angle scanning area below the trolley body, so that the points on the surfaces of the two rails form continuous strips in the point observation information, and then uses the texture related to the rail edges in the image observation information for mutual verification, so as to separate the points on the surfaces of the two rails and extract two rail top lines, and take the midline of the two rail top lines as the track center line, and determine the track reference plane with the surfaces of the two rails.

[0026] The data acquisition device then searches for linearly connected sparse points in the high-position slice of the point observation information along the direction of the line, verifies the continuity of the connected structure in the gray space by using the image observation information, takes the connected structure that passes the verification as the preliminary track of the wire, and forms a narrow band above and below the connected structure to obtain the wire reference band. In order to establish the position information of the support, the data acquisition device jointly detects columnar point clusters or strips extending vertically and having approximately constant cross sections in the point observation information and the intensity projection graph, and records the projection sequence of the columnar point clusters or strips on the track center line as a support calibration point column. In order to generate the intensity projection graph, the data acquisition device projects each point containing spatial position and reflection intensity to the pixel position on the imaging plane of the industrial camera according to the line-of-sight correspondence relationship between the three-dimensional laser radar and the industrial camera, and selects or fuses the projection points in the same pixel position according to time proximity and incident direction consistency to maintain the one-to-one alignment of the intensity projection graph and the image observation information, and reduces the ghosting near the wire reference band and the support calibration point column through boundary protection and occlusion suppression. The data acquisition device uses the trolley odometer to unify the travel mileage index of the above-mentioned various types of data, and outputs the track center line, the track reference plane, the wire reference band and the support calibration point column in each travel mileage interval, so that the point observation information, the image observation information and the intensity projection graph can be superimposed in the same coordinate system, forming the basis input required for the corridor reference graph.

[0027] The data processing device converts the point observation information, the image observation information and the intensity projection map into geometric hypotheses that can accumulate evidence in the parameter space under the same travel mileage index and the same coordinate convention, and provides spatial prior with the help of the corridor reference map to constrain and screen the hypotheses, so that the geometric configuration of the catenary locator, the dropper and the section insulator is stably separated and refined from the continuous observation in the form of peaks. Specifically, the data processing device first projects the three types of input onto the same reference plane by using the track center line and the track surface reference plane in the corridor reference map, so that the edge intermittent point traces from the point observation information, the edge pixels from the image observation information and the intensity jumps from the intensity projection map form a candidate edge set on the same reference plane; then the multi-domain Hough transform is performed under the mask constraint of the corridor reference map, that is, only the candidate edges consistent with or tangent to the wire reference band are allowed to vote in the line domain accumulation graph, only the candidate edges with curvature level compatible with the wire reference band neighbors are allowed to vote in the arc domain accumulation graph, and only the candidate edge pairs with stable intersection points or attachment relationship with the wire reference band are counted in the intersection domain accumulation graph, so that a large number of invalid hypotheses of redundant edges are excluded from the voting.

[0028] The peak column after multi-domain accumulation forms a wire reference axis along the travel direction, which corresponds to the wire reference band one by one and serves as a geometric carrier for subsequent relationship judgment; on this basis, relationship backfilling is performed, that is, the peaks supported by the wire reference axis vote reversely to the hypothesis space of their accompanying components according to the connection rules defined in the template dictionary: the short line peak of the catenary locator strengthens the independent intersection with the wire reference axis in the intersection domain accumulation graph, the nearly vertical peak column of the dropper strengthens the continuous hanging connection on both sides of the wire reference axis in the line domain accumulation graph, and the sequence group of the section insulator composed of alternating short line response and arc response is weighted between the arc domain accumulation graph and the line domain accumulation graph to consolidate its mode consistency. Since adjacent components may produce peak coupling in the same region, the data processing device establishes a discrete histogram for the direction of the candidate edge in each suspected coupling peak neighborhood, divides the angle scale into sub-direction clusters, and then independently reconstructs the source of the candidate edge in the point observation information, the image observation information and the intensity projection map, respectively, and selects the main peak and retains the secondary peak according to the connection consistency with the wire reference axis as the criterion to solve the short-time occlusion and overlapping ambiguity when the continuous slices are summarized. In order to improve the coarse-grained peak positioning to the engineering usable axis or contour, the data processing device performs micro-area Hough iteration around the peak in the three types of input by establishing a micro-area search window, and in the iteration, the new candidate edge is first limited by the category and direction of the current peak, then the local voting is performed in the corresponding Hough domain, and the center position and direction of the axis or contour are updated according to the voting, and the support edge set is shrunk to eliminate inconsistent edges, until the local geometric description tends to be stable in continuous iteration.

[0029] On the continuous slices along the track, the data processing device links the main peaks and the secondary peaks across the slices in the category consistency, the direction stability and the connection order stability to obtain the robust peak column; the robust peak column gives the positioning drop point and the stable pose in the corridor reference map after micro-area Hough iteration, wherein the positioning drop point of the catenary locator takes the intersection of the stable axis and the wire reference axis, the positioning drop point of the dropper takes the connection point of the stable axis and the wire reference axis, and the positioning drop point of the sectional insulator takes the center connection point of the stable contour on the wire reference axis. Based on the positioning drop point and the stable pose, the data processing device completes the geometric correlation with the track center line, the wire reference band and the support calibration point column in the corridor reference map, and gives the lateral offset value, the wire longitudinal position, the hanging point section number and the height value, and checks these quantities and the stable pose one by one in the category consistency, the position consistency and the connection consistency; the principle of the checking is to verify each other by using the high-confidence peak value formed by the prior limited domain Hough transformation and the connected structure formed by the relationship backfilling, so that the stability and the reviewable result are maintained in the presence of noise, shielding and uneven reflection.

[0030] Further, the data processing device projects each positioning drop point to the track center line in the track surface reference plane to obtain a mileage positioning point, takes the reading of the trolley mileage meter at the mileage positioning point as the longitudinal coordinate, and records the shortest horizontal distance projected to the track center line as the lateral offset value; the data processing device projects each positioning drop point to the wire reference axis along the direction perpendicular to the wire reference axis in the track surface reference plane to obtain a wire correlation point; takes the interval of the positioning drop point relative to the wire correlation point in the vertical direction as the height value, takes the sequential position of the positioning drop point along the wire reference axis as the wire longitudinal position, and finds the nearest two adjacent points in the support calibration point column, positions the mileage positioning point between the two points, and determines the hanging point section number in proportion to the distance to the two points.

[0031] Further, the point-like observation information and the image observation information are mutually verified under the same time reference and the same coordinate convention, so that the geometric significant structure presents stable characteristics of separability, traceability and projectability in the data domain. The point-like observation information is generated by the three-dimensional laser radar during the trolley traveling process, and the discrete three-dimensional points simultaneously carry the spatial position and the reflection intensity, and can form continuous sampling in the space where the track and the catenary are located; the image observation information is composed of a sequence of gray-scale images obtained by the industrial camera installed on the trolley, and has high contrast performance of texture and edge. By calibrating the relative pose of the three-dimensional laser radar and the industrial camera and performing time synchronization, the point-like observation information is projected onto the imaging plane of the industrial camera according to the line-of-sight relationship to obtain an intensity projection map, so that the intensity projection map and the image observation information are pixel by pixel aligned, thereby establishing the corresponding relationship between the point-like observation information and the image observation information at the pixel level.

[0032] In the structure extraction of the track region, the two groups of rail surface points in the point-like observation information have the performance of continuous flatness, normal consistency and relatively stable reflection intensity, so the two groups of rail surface points can be stripped from the mixed point cloud in the point-like observation information through geometric consistency and neighborhood continuity; the two approximately parallel rail top lines and the stable spatial relationship between the two approximately parallel rail top lines are used to eliminate the uncertainty caused by local occlusion and lack of sampling, and the geometric midlines have the property of symmetric minimum deviation in space, so that the midline of the two rail top lines is used as the track center line, and the track center line becomes the reference axis for subsequent projection and sorting; the plane defined by the two groups of rail surface points is naturally orthogonal to the track laying direction and keeps flat in a small scale, so the plane is used as the track surface reference plane to distinguish the sleepers, the ground and the overhead catenary elements in the vertical direction.

[0033] In the structure extraction of the catenary region, the point-like observation information is sliced according to the height above the track surface reference plane, and the high-level slicing can filter out the ground and vehicle interference in space and only keep the sparse point traces of the conductor and its adjacent components; the conductor shows the form of extending along the line and locally linear continuity in the high-level slice, and a penetrating point trace chain can be separated from the noise and isolated points by means of connectivity and direction consistency, the point trace chain remains stable and extends in the slice sequence, and thus the preliminary conductor trajectory extending along the line is obtained; the narrow band formed above and below the preliminary conductor trajectory with the preliminary conductor trajectory as the midline can meet two types of needs at the same time, one is to limit the search range of candidate edges and candidate points in space to reduce false detection, and the other is to provide a limited tolerance band for voting and registration in subsequent identification and positioning, and the narrow band is the conductor reference band.

[0034] In the extraction of the support structure, the columnar point cluster of the support column extends along the vertical direction and has an approximately constant cross section in the point-like observation information, and corresponds to the vertical stripe or vertical edge in the intensity projection graph and image observation information, and this cross-domain consistency enables the same morphological and connectivity criteria to be used for detection in the two types of observation; when the columnar point cluster or stripe independently detected in the point-like observation information and the intensity projection graph mutually corroborate each other in spatial position, it can be determined as a support column instance. Since the track center line provides a single-valued ordered reference along the line direction, the position of each support column in space is orthogonally projected onto the track center line, which can map the two-dimensional position into an ordered point sequence on the track center line, and the ordered point sequence naturally reflects the sequential relationship and spacing relationship of the support column along the line, so the projection sequence is recorded as the support column calibration point sequence.

[0035] Further, the data processing device comprises a candidate edge construction unit, a sub-domain Hough mapping unit, a conductor reference axis establishment unit, a relationship backfilling unit, a peak decoupling output unit and a result output unit; the data processing device unifies the point-like observation information and the image observation information in the projection alignment in the corridor reference map, takes the conductor reference band and the support calibration point column as the spatial prior, performs constraint evidence accumulation and relationship verification on the geometric hypothesis, and separates, connects and refines the catenary locator, the dropper and the sectional insulator in the parameter space in the form of peaks. The candidate edge construction unit first receives the point-like observation information and the image observation information, completes the pixel-level and slice-level projection alignment under the coordinate convention of the corridor reference map, uniformly expresses the edge pixels from the gray-scale image, the intensity jump edges from the intensity projection map and the boundary points from the point-like observation information slice, forms a candidate edge set carrying source marks and direction information, and the source marks of the candidate edge set are used for consistent weighting of different evidences in subsequent voting, thereby reducing the influence of single sensor distortion.

[0036] The sub-domain Hough mapping unit at least one of line domain, arc domain or intersection domain in the corridor reference map, generates a corresponding cumulative map; in the line domain, the candidate edge only participates in effective voting when it is consistent or tangent to the direction of the conductor reference band, so as to suppress the stray straight line response irrelevant to the conductor; in the arc domain, only the curvature level and direction combination compatible with the neighborhood of the conductor reference band are counted, so as to highlight the structural features of the sectional insulator and the like composed of short lines and arc segments; in the intersection domain, the intersection or attachment relationship between the candidate edge and the conductor reference band is recorded, so as to strengthen the local connection mode of the short line and the conductor such as the catenary locator; the three types of cumulative maps share the same advancing direction index and perform evidence merging at the same position, so that different sources form comparable vote numbers on the same parameter grid.

[0037] The wire reference axis establishing unit searches for a stable peak column in the cumulative graph along the travel direction, concatenates scattered local peaks into a wire reference axis through the connectivity and monotonicity of the peak column, and establishes a one-to-one correspondence with the wire reference band; the wire reference axis is used as a geometric skeleton to provide a unique attachment object for subsequent connection verification and position attribution, thereby constraining the candidate edges in free space to a narrow channel related to the wire. The relationship backfilling unit writes the connection relationship between the candidate and the wire reference axis and the support point column into the cumulative graph according to the template dictionary containing the catenary locator, dropper and sectional insulator, and triggers the accompanying part constraint: when the intersection domain appears a short line peak forming a stable intersection with the wire reference axis, the straight line hypothesis at this position is inversely weighted in the linear domain, forming a candidate for the catenary locator; when the linear domain appears a continuous peak column near vertical on both sides of the wire reference axis and spanning multiple slices, it is determined as a dropper candidate in combination with the dependent relationship of the support point column; when the linear domain and the arc domain present an alternating sequence of short lines and arc segments on the wire reference axis, the two domain tickets are mutually backfilled according to the connection order of the template dictionary, forming a candidate for the sectional insulator.

[0038] The peak decoupling output unit first determines the peak neighborhood in the cumulative graph for the peak coupling caused by adjacent components in the same area, then performs sub-direction cluster splitting according to the direction information of the candidate edge, divides the source edge into several sub-clusters according to the angle scale, and respectively traces the specific support of each sub-cluster in the point-like observation information, image observation information and intensity projection graph, and then independently reconstructs the candidate peak in each sub-cluster; each candidate peak is evaluated for connection consistency with the wire reference axis, which investigates the existence and stability of the intersection point, the relative side and dependent relationship, and the sequence retention along the wire reference axis, so as to select the main peak that best meets the connection rules and appears most stably in consecutive slices, and the remaining candidate peaks are retained as secondary peaks to resist the influence of short-time occlusion and intermittent sampling; then the main peak and the secondary peak are cross-slice summarized according to the category consistency, direction stability and connection sequence stability, forming a robust peak column with spatiotemporal continuous evidence.

[0039] The result output unit generates the positioning landing point and the stable pose in the corridor reference graph based on the robust peak column: the positioning landing point of the catenary locator is given by the intersection of the short line main peak and the wire reference axis, and the direction of the short line is taken as the directional component of the stable pose; the positioning landing point of the dropper is given by the connection point of the near-vertical main peak and the wire reference axis, and the direction of the main peak is taken as the directional component of the stable pose; the positioning landing point of the sectional insulator is given by the center connecting point of the alternating sequence of short lines and arc segments, and the main direction of the sequence and the trend of the arc segment together determine the stable pose; all positioning results are recorded under the coordinate convention of the corridor reference graph, so as to be geometrically related to the track center line, the wire reference band and the support point column, and enter the subsequent geometric parameter generation and consistency checking process.

[0040] The following provides an example of a template dictionary for three types of targets:

[0041] Catenary locator template dictionary: target geometric prototype is a short line element; spatial position constraint is that the short line center is located in the narrow band on both sides of the wire reference band, the projection of the short line length along the wire reference axis is a short range, and the short line normal direction is nearly vertical or limitedly inclined relative to the wire reference axis; connection rule is that there is a single intersection with the wire reference axis and the intersection remains in a monotonic order within the continuous slice, and the center line of the wire reference band is preferentially fitted; the dependent relationship with the support calibration point column is not mandatory, but the intersection density does not abnormally gather in the nearest neighbor section with the support calibration point column; Hough domain trigger is that the cross-domain cumulative graph first establishes a stable intersection, and then the short line peak is confirmed in the line domain cumulative graph; evidence backfill is that the cross-domain cumulative graph weights the same address grid of the line domain cumulative graph to suppress short line hypotheses that do not form an intersection with the wire reference axis; cross-slice consistency is that the category label and angle label are stable and the intersection order is not reversed; positioning drop point generation rule is to take the intersection of the stable short line peak and the wire reference axis as the positioning drop point, and take the short line direction as the direction component of the stable pose; conflict processing is that when there are multiple short line candidates in the same peak neighborhood, the connection consistency with the wire reference axis and the slice appearance stability are sorted, and the highest one is confirmed as the main peak, and the rest are kept as secondary peaks.

[0042] Suspender template dictionary: target geometric prototype is a peak column of nearly vertical line elements on continuous slices; spatial position constraint is that the line element and the wire reference axis have a unique connection point in the corridor reference graph, and the line element body is located in the feasible domain between the outer edge of the wire reference band on both sides and the track surface reference plane; connection rule is that the line element extends along the normal direction in the track surface reference plane with the connection point as the anchor point on the wire reference axis, and remains through or is short-time interpolatable in adjacent slices; dependent relationship with the support calibration point column is that the connection point is between the nearest neighbor support calibration point pair and consistent with the section number of the pair, and if it extends across the section, it is marked as abnormal for review; Hough domain trigger is that the line domain cumulative graph has nearly vertical strong peaks on both sides of the wire reference axis, and the cross-domain cumulative graph is used to assist in confirming the uniqueness of the connection point; evidence backfill is that the line domain cumulative graph weights the connection point grid of the cross-domain cumulative graph to strengthen the unique connection point, and performs time sequence smoothing on the same address grid of adjacent slices to enhance continuity; cross-slice consistency is that the category label is constant, the angle label fluctuation is limited, and the connection point order on the wire reference axis is not reversed; positioning drop point generation rule is to take the connection point of the nearly vertical main peak and the wire reference axis as the positioning drop point, and take the line element direction as the direction component of the stable pose; conflict processing is that when there are multiple nearly vertical candidates near the same position, the one that is consistent with the dependent relationship of the nearest neighbor support calibration point pair and has the highest cross-slice continuity is preferentially selected as the main peak.

[0043] The segmented insulator template dictionary: the target geometric prototype is an alternating sequence group of short line responses and arc segment responses on the conductor reference axis; the spatial position constraint is that the whole sequence group falls within the conductor reference band, the short line responses and arc segment responses are alternately arranged around the conductor reference axis, and the deviation of the sequence center from the conductor reference axis is limited; the connection rule is that the sequence head and tail are continuous with the conductor reference axis, and the appearance order of the internal short line and arc segment is consistent with the template definition; the dependent relationship with the column of support calibration points is not mandatory, but the sequence should be located in the sliding window between the adjacent support calibration point pairs; the Hough domain trigger provides short line response candidates for the line-shaped domain cumulative graph and arc segment response candidates for the arc-shaped domain cumulative graph, and the two domains present an alternating peak column in the same travel mileage interval of the conductor reference axis; the evidence backfilling is that the line-shaped domain cumulative graph and the arc-shaped domain cumulative graph are weighted with respect to each other in the same address cell, only combinations satisfying the alternating relationship are retained, and the attachment or intersection of the short line and the conductor reference axis is verified on the cross-domain cumulative graph to exclude false arc segments; the cross-slice consistency is that the category label is constant, the relative order of the short line and the arc segment remains consistent in the continuous slices, and the sequence center monotonically advances in the travel direction; the positioning landing point generation rule is to calculate the center connection point of the sequence group on the conductor reference axis as the positioning landing point, and to take the sequence main body direction and the arc segment trend to determine the direction component and the contour description of the stable pose; the conflict processing is that when two sequence groups partially overlap and share short line responses, the sequence with higher joint votes of the line-shaped domain cumulative graph and the arc-shaped domain cumulative graph is selected as the main sequence, and the other sequence is marked as a secondary sequence for subsequent time sequence verification decision.

[0044] Further, the candidate edge construction unit first aligns the point-like observation information, the image observation information and the intensity projection graph under the unified time reference and the corridor reference graph plane, so that data from different sources are comparable, superimposable and traceable in the same reference. The purpose of edge detection on each grayscale image and the corresponding intensity projection graph in the image observation information is to extract the intensity jump position most sensitive to the geometric contour in the texture domain and the reflection domain at the same time. The edge detection output is collected as an edge pixel set, and the source label and direction information of each pixel are retained for evidence differentiation and consistency check during subsequent voting. At the same time, the point-like observation information is sliced according to the height above the track surface reference plane and projected to the corridor reference graph plane piece by piece. The jump boundary of point density and reflection intensity is extracted by measuring the neighborhood difference of the slice grid, and the slice boundary point set is obtained.

[0045] The neighborhood difference method uses the fact that the medium change from solid to background of the overhead contact line and the support will cause the sampling distribution and echo intensity to change abruptly, so it can stably expose the boundary of the geometric profile in the point cloud slice. To avoid systematic bias introduced by a single source under conditions of occlusion, uneven reflection or short-term distortion, the candidate edge construction unit combines the edge pixel set and the slice boundary point set into a candidate edge set, and during merging, it performs registration, deduplication and consistency check on multiple source evidence at the same location, retains edges that can be mutually confirmed from both gray scale and intensity perspectives, and uses source markers and direction information as priori for subsequent domain Hough mapping. The domain Hough mapping unit performs line domain Hough mapping on the candidate edge set in the corridor reference map, establishes a line domain accumulation graph containing only angle scale and distance scale, and uses the guide line reference band as a mask, so that only the candidate edges consistent or tangent to the direction of the guide line reference band are allowed to vote. The mechanism of this mask is to exclude a large number of irrelevant straight line hypotheses from the parameter space, so that the accumulation energy is concentrated on the solution compatible with the guide line direction, thereby quickly forming a peak column continuous along the direction of travel and providing high signal-to-noise ratio evidence for the establishment of the guide line reference axis.

[0046] In addition to the linear structure, there are also short elements in the catenary system that are in the form of arcs or bends. Therefore, the candidate edge set in the same region is subjected to arc domain Hough mapping, and an arc domain accumulation map containing angle scale and curvature level is established. The arc domain accumulation map accumulates the candidate edges in the angle and curvature parameter plane, converges the local near-arc, bend or sleeve transition edges into stable peaks, and relies on the spatial constraints of the conductor reference strip neighborhood to suppress the false triggering of background arc texture, thereby highlighting the true response of the insulator shape. For catenary locators and other targets characterized by short line attachment or intersection, simple linear or arc domain accumulation cannot directly distinguish them. Therefore, cross domain Hough mapping is introduced to establish a cross domain accumulation map that records the intersection relationship of two line segments. The principle is to combine the candidate edges in pairs and judge whether they form an intersection point or a stable attachment relationship on the corridor reference map plane. Only the combination that forms a single intersection point or short distance attachment with the conductor reference strip is recorded as a valid vote, thereby depositing the high-level relationship of "topology connection between short line and conductor" into the peak value in the parameter space. The three types of accumulation maps share the same coordinate convention and the same progress distance index. The source markers provided by the candidate edge construction unit are preserved in the domain Hough mapping unit and used for multi-source evidence merging and consistency screening on the same parameter grid, avoiding accidental peaks in a single domain dominating the decision. The conductor reference strip and the support calibration point serve as spatial priors throughout the domain Hough mapping process, which is reflected in three aspects: first, mask the spatial range of voting to reduce the chance of candidate edges entering the parameter space; second, weight the candidate edges that meet the membership and order relationship during voting to promote the true structure to accumulate into a distinguishable peak faster; third, provide a peak column connectivity and monotonicity test benchmark in the accumulation map post-processing stage, so that the true structure along the progress direction appears as a continuous, smooth and monotonically advancing energy ridge line in the parameter space.

[0047] Further, the relationship backfilling unit converts the dispersion peak value obtained by the sub-domain Hough mapping into a structured evidence chain conforming to the connection rule by using the geometric main shaft provided by the wire reference axis and the narrow band defined by the wire reference band. First, the short line peak in the cross-domain cumulative graph is searched in the narrow band on both sides of the wire reference axis. The intersection or tangency of the short line peak and the wire reference axis means that there is a local contact relationship of "short line and wire" on the corridor reference graph plane. Such a relationship is a direct representation of the geometric configuration of the catenary locator. The position and direction of these short line peaks are backfilled to the line domain cumulative graph as anchor points. The principle is to translate the topological evidence of "local connection" into geometric evidence of "straight line hypothesis". The short line hypothesis consistent or tangent to the direction of the wire reference band obtains secondary accumulation in the parameter space, thereby suppressing false straight line responses caused by texture or noise without real connection with the wire reference axis. When the short line peak shows a stable order along the travel direction on the wire reference axis, it means that these intersection or tangency events maintain a monotonic progression rather than random jumps in adjacent slices. The stable order and the one-to-one correspondence of the mileage together constitute the basis for the time consistency and spatial consistency of the catenary locator, so as to confirm the corresponding locator candidate.

[0048] Subsequently, starting from each candidate connection point on the wire reference axis, a nearly vertical peak column is searched in the line domain cumulative graph along the normal direction in the rail plane reference plane. The nearly vertical peak column reflects the existence of a continuous downward linear structure at the candidate connection point, which is consistent with the configuration of the dropper hanging downward from the wire. In order to avoid misidentifying accidental vertical texture as a real component, it is necessary to verify the continuity of the peak column in adjacent slices. Only the peak column that maintains a stable direction, a smooth position progression and a unique connection point with the wire reference axis in adjacent slices is considered as a dropper candidate that meets the geometric and topological double constraints. For segmented insulators, a single domain peak value is insufficient to express the complex feature of "alternating short lines and arcs". Therefore, a sliding window is set along the wire reference axis to search for a sequence group composed of multiple short line peaks and an arc-shaped response in the line domain cumulative graph and the arc domain cumulative graph. The function of the sliding window is to limit the local context, so that the short line peak and the arc-shaped response must appear in the same mileage interval and form an alternating pattern around the wire reference axis. When the sequence group meets the conditions of being located on the wire reference axis, the short line peak and the arc-shaped response appearing alternately in the travel direction, and the connection order being consistent with the template dictionary, it can be considered that it not only geometrically exists, but also semantically conforms to the definition of the segmented insulator, so as to confirm the segmented insulator candidate.

[0049] The three branches operate under the same coordinate system. The backfilling vote of short-line peaks strengthens the connectable positions in the cumulative graph of the linear domain. The near-vertical search downwards of candidate connection points extends straight-line evidence into through-connection evidence. The alternating verification of the linear and arc domains elevates single-peak evidence into sequence evidence. Together, these three methods reduce sporadic peaks caused by noise, occlusion, and uneven reflection. Finally, the locator candidates, dropper candidates, and segmented insulator candidates are merged into a candidate set of three types of targets under the same mileage index. During the merging, the position on the conductor reference axis, the side difference from the conductor reference strip, and the continuity in adjacent slices are used as consistency criteria. Overlapping candidates are deduplicated and consistency-screened, retaining candidates with clear connection relationships, complete sequence patterns, and cross-domain evidence that corroborates each other. Through this process of triggering the cross-domain cumulative graph, backfilling the linear and arc-shaped cumulative graphs, and then backfilling the relationship organized by the conductor reference axis, the discrete peaks in the parameter space are reorganized into structured assumptions consistent with the actual assembly relationship of the contact network. This enables the contact network locators, droppers, and segment insulators to be identified in a verifiable and traceable manner within the neighborhood of the conductor reference axis, providing high-confidence input for subsequent peak decoupling and robust peak generation.

[0050] Furthermore, the execution process of the peak decoupling output unit includes: detecting the highest intensity peak in the cumulative graph and determining whether there are multiple source edges spatially overlapping or adjacent to it, forming a clustered response; when peak coupling is determined, expanding from high to low cumulative intensity along the peak as the center until the cumulative intensity drops to a preset neighborhood threshold or a local intensity valley is encountered, the resulting closed region is defined as the peak neighborhood, and the source edges within the peak neighborhood are extracted as the objects to be split. Within the peak neighborhood, the direction of each source edge is calculated and discretized according to a preset angle scale, and source edges whose directions fall into the same angle scale are divided into the same sub-cluster; when the number of source edges at the angle scale boundary is insufficient to form stable support, they are merged into the sub-cluster corresponding to the adjacent angle scale according to the nearest neighbor angle principle; after the division is completed, several sub-clusters are obtained, each sub-cluster carrying its angle scale label and the corresponding set of source edges.

[0051] For each sub-cluster, the specific locations of the source edges in the original observation are traced back, the source edges including the edge pixels from the image observation information, the edge pixels from the intensity projection map and the slice boundary points from the point-like observation information slice; in the tracing back process, the isolated edges and broken edges irrelevant to the peak neighborhood are removed, and the source edges having continuous support relationship with the peak neighborhood are reserved; then, only taking the source edges of the sub-cluster as input, the Hough mapping is independently performed in the Hough domain corresponding to the angle scale of the sub-cluster to generate the candidate peaks of the sub-cluster, and a plurality of candidate peaks and their response positions in the cumulative map are obtained. For each candidate peak, the connection consistency with the reference axis of the guide line is evaluated, the connection consistency including three contents, respectively, whether the intersection with the reference axis of the guide line exists and is stable, whether the relative positions of the candidate peaks on both sides of the reference axis of the guide line conform to the subordinative relationship defined by the template dictionary, and whether the order of the candidate peaks on the reference axis of the guide line is consistent with the order of the adjacent components; in the same peak neighborhood, the candidate peak with the highest connection consistency and the most stable appearance in the continuous slices is selected as the main peak, and the remaining candidate peaks are reserved as the secondary peaks for subsequent use; when there are a plurality of candidate peaks with the same connection consistency and stability, the one with smaller distance from the reference axis of the guide line is selected as the main peak.

[0052] The continuous slice sequence is obtained along the line direction, and the positions of the main peak and the secondary peak, the angle scale label and the category label are recorded in each slice; the cross-slice summarization is performed between the continuous slices according to the following three rules, respectively, the category consistency, the direction stability and the connection order stability; the category consistency means that the peak category labels of adjacent slices remain consistent or consistent within the allowed template equivalent range, the direction stability means that the angle scale labels of adjacent slices change by no more than a preset angle tolerance band, and the connection order stability means that the order of the peaks on the reference axis of the guide line does not occur in reverse order; when any rule is broken and a short interruption occurs, the interpolation is allowed to be performed through the main peak and the secondary peak of the adjacent slices within a limited slice interval to maintain continuity; finally, the same type of peaks satisfying the above three rules are linked as robust peak columns on the continuous slices. Taking each robust peak column as the center, a micro area search window is established in the original observation, and the source edges in the micro area are extracted as the support edge set; according to the category label of the robust peak column, the corresponding Hough domain is selected for local Hough mapping in the micro area, and the local peak is obtained and the geometric description of the robust peak column is updated according to the local peak, the geometric description including the center position, direction and boundary range of the axis or contour; the micro area search window is shrunk and the source edges not conforming to the current geometric description are removed according to the updated geometric description, and the remaining source edges form a new support edge set; the steps of local Hough mapping, geometric description updating and support edge set shrinking are repeated until the change of the geometric description updated continuously two times is lower than a preset convergence threshold or a preset maximum iteration number is reached, and a stable axis or a stable contour is obtained.

[0053] Based on the stable axis or stable profile obtained through micro-area Hough iteration of the robust peak array, the positioning points and stable poses of the corresponding targets are generated within the corridor reference reference. Specifically, the positioning point of the contact wire locator is the intersection of the stable axis and the conductor reference axis, and the stable pose is the center position and direction of the stable axis in the corridor reference reference. The positioning point of the dropper is the connection point of the stable axis and the conductor reference axis, and the stable pose is the center position and direction of the stable axis in the corridor reference reference. The positioning point of the segmented insulator is the center connection point of the stable profile, composed of alternating short-line and arc-shaped responses, on the conductor reference axis, and the stable pose is the center position, main orientation, and contour boundary range of the stable profile in the corridor reference reference. All positioning points and stable poses are recorded using the corridor reference reference as the coordinate system and are correlated with subsequent geometric parameter calculations and consistency checks.

[0054] The following is given an example of a straight line segment with a length of... (in This is a complete example of an intelligent inspection vehicle integrating contact wire 1C and 4C on a railway line (where the inspection section length is in meters). Assume the vehicle travels at a constant speed. (in (This indicates the speed of the vehicle, in meters per second.) The 3D lidar is installed approximately [distance] from the reference plane of the track surface. (in (This indicates the installation height of the 3D LiDAR, in meters). The industrial camera is installed approximately [distance] from the reference plane of the rail surface. (in (This indicates the installation height of the industrial camera, in meters), and the sampling frequency of the point observation information is... (in The sampling frequency of the 3D lidar (in Hertz) represents the frame rate of the image observation information. (in This refers to the frame rate of an industrial camera, measured in Hertz (Hz).

[0055] First, establish the projection relationship between the camera and the 3D LiDAR. Let the i-th coordinate system be the coordinate system of the 3D LiDAR. The points are (in The integer number representing the point index. , , Let these represent the three axial coordinates of the point in the 3D LiDAR coordinate system (in meters). Assume the extrinsic parameters of the 3D LiDAR to the industrial camera are... (in To convert from the 3D LiDAR coordinate system to the industrial camera coordinate system Rotation matrix For 3D LiDAR to industrial cameras Translation vector, unit: meter), let the intrinsic matrix of the industrial camera be (Wherein is the camera intrinsic matrix, the element unit is pixel).

[0056] The visual calibration board and synchronous trigger are used to obtain (Wherein represents the number of synchronous frame pairs used for calibration, unit: frame pair) group corresponding frames, and a group of values are obtained after minimizing the re-projection error: , (Wherein , , respectively represent the right-hand rule rotation around the camera coordinate system , , axis, angle unit: degree), (unit: meter). Any point projected onto the imaging plane of the industrial camera has pixel coordinates (Wherein , respectively represent the horizontal and vertical pixel coordinates, unit: pixel) calculated as follows: , wherein is the scale factor of homogeneous projection (dimensionless). Assign the reflection intensity of each to the corresponding position to obtain the intensity projection map, and align the intensity projection map with the image observation information at the same time through timestamp alignment.

[0057] To determine the rail surface reference plane and the track center line, two groups of steel rail upper surface points are extracted from the point-like observation information and plane fitting is performed. Let the rail surface be ; (Wherein , , are the three-axis coordinates of any point in the three-dimensional laser radar coordinate system, unit: meter; , , , are plane coefficients, , , , together constitute the normal, is a constant term, all of which are dimensionless normalized coefficients), the least squares solution is used on the rail top candidate points within the distance threshold (Wherein is the maximum allowed distance from the point to the plane for plane fitting, unit: meter) to obtain the normal vector is the unit normal vector, dimensionless) and unitize. Slice the point cloud by height, slice thickness where is the slice thickness, in meters, and compute local height gradient within each slice with neighborhood radius where is the neighborhood search radius, in meters. Separate two approximately parallel rail top lines with threshold where is the height gradient threshold, in meters. Fit two rail top centerlines with piecewise polyline respectively and where , denote the polyline curve composed of ordered vertex set, in meters. Generate the rail centerline where is the midline polyline of the two rail top polylines, in meters. Define the corridor reference frame with normal and together, and denote the milepost coordinate along as where is the milepost parameter, in meters.

[0058] To obtain the traverse reference strip, slice the point-like observation information in the region where is the lower height threshold of the high slice, in meters. Slice with slice thickness where is the high slice thickness, in meters. Concatenate linear point traces with connected length where is the minimum connected length of linear point traces, in meters, and directional standard deviation in degrees to obtain the preliminary traverse trajectory where is the preliminary polyline trajectory of the traverse, in meters. Generate the traverse reference strip with width and outer expansion width where , are the inner and outer expansion widths, in meters, respectively, in the corridor reference frame plane with inner expansion width where is the strip width, in meters, to obtain the traverse reference strip where is the traverse reference strip region. The column calibration point list is obtained by joint detection of point-like observation information and intensity projection map: in each milepost slice, if the columnar connected height along the rail surface reference plane normal direction and the cross-section variation rate (both in meters and dimensionless ratio), then orthogonally project them to to get the calibration points, arrange them in ascending order to form (where represents the projection mileage point of the th pillar on the th image, in meters; here represents the number of calibration points, which is a dimensionless integer). This paragraph obtains calibration points, with an average distance of about meters.

[0059] In the candidate edge construction stage, the Canny edge is used for the gray image, with low and high thresholds (where , are the gradient thresholds, in pixel gray levels) to obtain the edge set (where is the edge pixel set of the gray image, is the th edge pixel), with each pixel attached to the gradient direction (where is the edge direction angle, in degrees); the intensity projection map is also extracted (where is the edge pixel set of the intensity projection map); a grid is established on the corridor reference map for the point-like observation information, with a grid side length of (where is the grid size, in meters), and the grid neighborhood difference is calculated to obtain the slice boundary point set (where is the boundary point set of the point cloud projection). The three source sets are fused into the candidate edge set (where is the candidate edge set), with the source markers and directions retained (where is the direction angle of the candidate edge, in degrees).

[0060] The multi-domain Hough mapping adopts three-domain accumulation. The linear domain adopts parameterization (where is the signed distance to the origin, in meters; is the normal angle, in degrees), and the voting of the candidate point (where , are the corridor reference map plane coordinates, in meters) to satisfies . The linear domain accumulation map is established (where ​is the voting accumulator of the line domain parameter plane, in votes, and the guide line reference band mask where is a binary function indicating whether a plane point is located within the guide line reference band or outside the guide line reference band , respectively, the voting is allowed only when the corresponding direction makes an angle with the tangent of the guide line center line in degrees or is tangent to the guide line center line. The arc domain is parameterized by the angle and the curvature level where is the quantized form of the curvature, in reciprocal meters, the edge points of the local arc segment are fitted to a sub-circle by least square method where , are the coordinates of the sub-circle center, in meters, and is the radius of the sub-circle, in meters, the curvature is quantized to a discrete level, and accumulated to where is the voting accumulator of the arc domain. The intersection domain records the intersection relationship between the short line and the guide line: for two lines, assume and where is the signed distance between the two lines, in meters, and is the normal angle between the two lines, in degrees, if the intersection point where , are the coordinates of the intersection point on the plane, in meters, satisfies and the normal distance to the guide line center line is less than where is the maximum allowed normal distance from the intersection point to the guide line center, in meters, then the intersection domain accumulates where is the voting accumulator of the intersection domain, and records the intersection point at . Numerical settings: line domain parameter step size , where is the angle quantization step size, in degrees, and is the distance quantization step size, in meters, arc domain , where is the angle step size of the arc domain, in degrees, and is the number of curvature levels, dimensionless. Along the mileage perform peak column search on using dynamic programming in bandwidth The energy ridge is located internally to obtain the reference axis of the wire. (in (This refers to the traverse parameter axis along the mileage direction, in meters) and... One-to-one correspondence.

[0061] Relationship backfilling is performed according to three categories of objectives. Contact wire locator: in Within the narrow bands on both sides, from Reading and List of intersecting short-term peaks (in For the first A short-term peak, (integer index, dimensionless), for each exist Co-location grid weighted backfill (in To fill in the weights (unit: votes), and check the monotonicity of the intersection sequence according to the mileage sequence, continuous (in If the slice remains monotonic (for slice counting, dimensionless), it is confirmed as a candidate for locator. Suspension string: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Starting from each candidate intersection point of the locator, in Internal search in near vertical direction The peak sequence (with interval endpoints in degrees) requires a cross-section length of [missing information]. (Unit: meters) If the center of the slice drifts within an adjacent slice (Unit: meters) If a unique connection point is established, it is confirmed as a candidate for dropper. Segmented insulator: In Set the length of the sliding window above. (in (Window length, in meters) Statistics within the window short-term peak sequence (in For the first A short-term peak, (integer index) and arc peak (in For the first A peak in an arc segment (using integer indices) to construct alternating scores ,in Scoring for alternating patterns, dimensionless; , , These are weighted coefficients, dimensionless; The number of times the "alternating short lines and arcs" rule is violated, dimensionless; take .when (in Threshold for segment insulator, dimensionless, and sequence center to Normal deviation Confirmed as segment insulator candidate when the distance is less than

[0062] Peak coupling caused by adjacent components, within each cumulative peak neighborhood radius Wherein is the neighborhood radius on the cumulative plot parameter grid, unit: grid Wherein is the angular step of sub-direction cluster, unit: degree , , Independent reconstruction of peaks after excluding inconsistent sources, calculate the connection consistency score with Wherein is the total score of connection consistency, dimensionless; is the intersection stability score, dimensionless; is the side consistency score, dimensionless; is the mileage sequence consistency score, dimensionless; is the corresponding weight, dimensionless, taking The highest is the main peak, and the rest are secondary peaks. For each type of candidate, a microzone window size Wherein is the microzone window edge length, unit: meter, is executed. Set the line parameters of the th iteration as or the arc segment parameters as Wherein is the non-negative integer iteration index, dimensionless, only keep the support edge set consistent with or in the window, and recalculate or Stop when and Wherein is the distance convergence threshold, unit: meter; is the angle convergence threshold, unit: degree, or the maximum iteration number Wherein is the maximum iteration number, dimensionless integer, to get the stable axis or stable contour.

[0063] Positioning drop point calculation: the intersection of the stable short line and Wherein , , ​The positioning drop point is taken as the coordinate of the drop point on the reference plane of the corridor, with the unit of meter. The hanging point is taken as the connection point of the stable near-vertical line and . The sectional insulator is taken as the stable sequence center connection point. The positioning drop point is orthogonally projected to to obtain the mileage coordinate , the lateral offset value (wherein is the shortest horizontal distance from the positioning drop point to , with the unit of meter), the height value (wherein is the vertical distance of the positioning drop point relative to the reference plane of the track surface, with the unit of meter), and the conductor longitudinal position is the sequence position along . The nearest two calibration points in the support post calibration point column are , (with the unit of meter), and thus the hanging point section number is ; (wherein is the continuous number of the hanging point section, dimensionless; is the left-adjacent support post index, dimensionless integer; is the mileage position of the adjacent support post, with the unit of meter; here is the mileage coordinate of the target, with the unit of meter). Within the window of , the linear domain short line peak and the arc domain arc segment peak appear alternately, the cumulative score is , the cumulative score is , the violation number is , and thus , the sectional insulator is confirmed, the sequence center mileage is , the normal deviation is (with the unit of meter), the positioning drop point thereof is taken as the center connection point on , and the height value is calculated as according to the center height of the stable contour. Finally, the height value , the lateral offset value , the conductor longitudinal position, and the hanging point section number of each target are checked for category consistency, position consistency, and connection consistency with the stable pose thereof, and the passers are written into the geometric parameter result set, and the non-passers are registered as the abnormal record set; in the above example, the catenary positioner at and the dropper at are consistent, and the isolated short line detected at is registered as an abnormal record because there is no stable intersection point with .

[0064] While specific embodiments of the application have been described above, it will be appreciated that those skilled in the art within the scope of the application can make modifications, substitutions and changes in the methods and systems described above without departing from the spirit and essential characteristics of the application. For example, acts recited in the foregoing method steps in combination, in the same or an equivalent method, for carrying out the same functions to achieve the same results while maintaining the essential characteristic of that which is recited are within the scope of this application. Accordingly, the scope of the application is to be construed in accordance with the appended claims rather than read expansively with reference to the description.

Claims

1. A smart inspection vehicle integrating 1C and 4C of overhead contact lines, characterized in that, include: The system comprises a trolley body, a data acquisition device, and a data processing device. The data acquisition device, while the trolley body is moving, simultaneously acquires point-based and image-based observation information along the contact wire corridor, as well as an intensity projection map obtained by projecting the point-based observation information through a line of sight and aligning each pixel with the image-based observation information. Based on the point-based observation information, it determines the track surface reference plane; based on the image-based observation information, it determines the traverse reference zone; and based on both point-based and image-based observation information, it determines the support column calibration point series. The track centerline, track surface reference plane, traverse reference zone, and support column calibration point series together constitute a unified corridor reference diagram. The data processing device uses the corridor reference diagram as a guide for... The domain Hough transform and relation backfilling algorithm identifies and locates contact wire locators, droppers, and segment insulators under the constraints of a corridor reference diagram. It ultimately generates the corresponding target's positioning point and stable pose within the corridor reference diagram, and calculates the lateral offset value, conductor longitudinal position, and suspension point segment number based on the positioning point. The algorithm then checks the height value, lateral offset value, conductor longitudinal position, suspension point segment number, and stable pose against category consistency, position consistency, and connection consistency. If they match, the results are written into the geometric parameter result set; otherwise, they are registered as anomaly records. The data processing device includes: a candidate edge construction unit, a domain Hough mapping unit, and a conductor reference axis establishment unit. The system comprises a relation backfilling unit, a peak decoupling output unit, and a result output unit. The candidate edge construction unit takes point observation information and image observation information as input, extracts edges, and projects them for alignment to form a candidate edge set for subsequent voting. The domain-specific Hough mapping unit performs Hough mapping on the candidate edge set in at least one Hough domain, generating at least one cumulative graph. The Hough domain includes any one or more of linear, arc, or cross domains. The entire Hough mapping process is spatially constrained by the traverse reference strip and the support calibration point column, allowing only compatible candidate edges to generate valid votes. The traverse reference axis establishment unit searches for and connects stable peak columns along the direction of travel in the cumulative graph. The reference axis of the conductor is used as a reference axis, which corresponds one-to-one with the reference strip of the conductor and serves as the object to which subsequent relationship judgment and positioning are attached. The relationship backfilling unit writes the connection relationship between the candidate and the reference axis of the conductor and the column calibration point column into the cumulative graph based on the template dictionary containing the contact wire locator, dropper and segment insulator. It performs relationship backfilling and accompanying component constraints to obtain the candidate set of three types of targets. The peak decoupling output unit performs peak decoupling processing on the peaks that are coupled due to adjacent components, and performs cross-slice summarization on the continuous slice according to the consistency of category, direction stability and connection sequence stability to form a robust peak column, and outputs the positioning landing point and stable pose of each target.

2. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 1, characterized in that, The data processing device projects each positioning point onto the track centerline in the track reference plane to obtain the mileage positioning point. The reading of the trolley odometer at the mileage positioning point is used as the longitudinal coordinate, and the shortest horizontal distance projected to the track centerline is recorded as the lateral offset value. Each positioning point is projected onto the guide reference axis in the direction perpendicular to the guide reference axis in the track reference plane to obtain the guide association point. The vertical distance between the positioning point and the guide association point is read as the height value. The sequential position of the positioning point along the guide reference axis is read as the longitudinal position of the guide. The two nearest adjacent calibration points are found in the support calibration point column, and the mileage positioning point is located between the two calibration points. The hanging point section number is determined by the ratio of the distance to the two points.

3. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 2, characterized in that, The point observation information acquired by the data acquisition device is a discrete set of three-dimensional points output by the three-dimensional lidar during the movement of the trolley. Each point contains its spatial location and reflection intensity. The image observation information is a sequence of grayscale images acquired by an industrial camera mounted on the trolley in the direction of travel. The points on the upper surfaces of the two rails are segmented from the point observation information, and two approximately parallel rail top lines are extracted. The midline of the two rail top lines is taken as the centerline of the track, and the plane jointly determined by the upper surfaces of the two rails is taken as the reference plane of the track surface. In the high-level slice of the point observation information, the preliminary traverse trajectory extending along the line is obtained based on the continuity of the linear features. A narrow band is formed above and below the preliminary traverse trajectory as the centerline to obtain the traverse reference band. The columnar point clusters or stripes extending vertically and with approximately constant cross-sections are detected in the point observation information and intensity projection map, and the projection sequence formed by them on the centerline of the track is recorded as the support calibration point sequence.

4. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 3, characterized in that, The specific process executed by the candidate edge construction unit includes: performing edge detection on each frame of grayscale image and the corresponding intensity projection map in the image observation information to obtain the edge pixel set; slicing the point observation information at height intervals and projecting it onto the corridor reference plane, extracting the jump boundary using the neighborhood difference method to obtain the slice boundary point set; and combining the edge pixel set and the slice boundary point set into the candidate edge set.

5. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 4, characterized in that, The execution process of the domain Hough mapping unit includes: performing linear domain Hough mapping on the candidate edge set within the corridor reference to establish a linear domain cumulative map containing only angle and distance scales, and using the conductor reference strip as a mask to allow only candidate edges that are aligned with or tangent to the conductor reference strip to vote; performing arc domain Hough mapping on the candidate edge set in the same area to establish an arc domain cumulative map containing angle scales and curvature levels to respond to short elements that are arc-shaped or bent in the insulator shape; and performing cross domain Hough mapping on the candidate edge set to establish a cross domain cumulative map that records the intersection relationship of two line segments to detect short line elements that intersect or are adjacent to the conductor.

6. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 5, characterized in that, The execution process of the relation backfilling unit includes: within the narrow bands on both sides of the conductor reference axis, using the cross-domain cumulative graph to find short peaks that intersect or are tangent to the conductor reference axis, and backfilling votes to the linear domain cumulative graph; when the intersection points of the short peaks show a stable order on the conductor reference axis, the corresponding locator candidate is confirmed; starting from each candidate connection point on the conductor reference axis, searching for near-vertical peak sequences along the normal direction in the rail reference plane in the linear domain cumulative graph, and verifying their continuity in adjacent slices; the verified peak sequences are used as dropper candidates; segment insulator relation backfilling: within the sliding window along the conductor reference axis, combining the linear domain cumulative graph and the arc domain cumulative graph, searching for a sequence group composed of multiple short peaks and an arc response; when the sequence group satisfies the conditions of being located on the conductor reference axis, short peaks and arc responses appearing alternately, and the connection order being consistent with the template dictionary, it is confirmed as a segment insulator candidate; locator candidates, dropper candidates, and segment insulator candidates are merged to form a candidate set of three types of targets.

7. The integrated intelligent inspection vehicle for overhead contact lines 1C and 4C as described in claim 6, characterized in that, The execution process of the peak decoupling output unit includes: for peak coupling caused by adjacent components in the same region, performing sub-direction cluster splitting: dividing the source edge into multiple sub-clusters according to the angle scale in the peak neighborhood, reconstructing them independently after tracing back their source, and selecting the main peak based on the connection consistency with the conductor reference axis, while retaining the rest as secondary peaks; on continuous slices along the line, summarizing each type of candidate according to category consistency, directional stability, and connection order stability to form a robust peak series; performing micro-area Hough iteration on the robust peak series in the original observation, gradually reducing the support edge set and refining the axis or contour, and finally generating the positioning landing point and stable pose of the corresponding target in the corridor reference; the original observation is a joint data set formed by aligning point observation information, image observation information, and intensity projection map at the same travel time.

Images