An underground cavity automatic detection method and system based on image recognition

Abstract

The application discloses an underground cavity automatic detection method and system based on image recognition, which comprises the following steps: collecting data for correction processing to generate a multi-frequency underground detection image group; recognizing continuous reflection ridge lines in the multi-frequency underground detection image group, connecting to generate cross-frequency homologous reflection tracks and track representation data; performing curved surface fitting to generate a propagation time reference surface; calculating a difference to generate a propagation time difference sequence, recognizing a local propagation time advance segment and determining a propagation time advance amount; connecting the local propagation time advance segment to generate a cross-frequency propagation time difference gap chain and gap chain space representation data; mapping the gap equivalent thickness to an underground space, and performing superposition and boundary closure to generate a propagation time difference gap closed body; and generating an underground cavity region and determining the position, burial depth and spatial range of the underground cavity. The application utilizes multi-frequency echo images and propagation time difference inversion to realize accurate positioning of the underground cavity, complete boundaries and strong anti-interference.

Description

Technical Field

[0001] This invention relates to the field of automatic underground cavity detection technology, and in particular to an automatic underground cavity detection method and system based on image recognition. Background Technology

[0002] Currently, underground cavity detection typically uses ground-penetrating radar to collect underground echo data along the detection path, and identifies abnormal areas by changes in echo amplitude, phase, hyperbolic characteristics, or interface reflection. Some methods combine multi-frequency detection images, propagation velocity, and echo arrival time to estimate the location and depth of underground anomalies.

[0003] However, current methods often rely on single-frequency band or local echo characteristics as the basis for judgment, making it difficult to correlate continuous reflection information from the same underground medium interface across different frequency bands. When the cavity reflection is weak, the medium interface is undulating, or there is clutter interference, it is easy to miss or falsely detect cavities, and it is difficult to use the propagation time variation to invert the cavity thickness in advance and determine the complete spatial boundary. Summary of the Invention

[0004] One objective of this invention is to propose an automatic detection method and system for underground cavities based on image recognition. This invention utilizes multi-frequency echo images and propagation time difference inversion to achieve accurate location of underground cavities with intact boundaries and strong anti-interference capabilities.

[0005] An automatic detection method for underground cavities based on image recognition according to an embodiment of the present invention includes the following steps: Multiple continuous detection paths are deployed along the underground area to be detected to determine the detection location. Data is collected using multiple detection frequency bands and corrected to generate a multi-frequency underground detection image group. Identify continuous reflection ridges in multi-frequency underground detection image groups, connect continuous reflection ridges with consistent spatial location, depth arrangement and echo response under different detection frequency bands, and generate cross-frequency homogeneous reflection trajectories and trajectory characterization data; Based on trajectory characterization data, surface fitting is performed on the overall trend of echo arrival time of each cross-frequency co-source reflection trajectory to generate propagation time reference surfaces corresponding to each underground medium interface. The difference between the actual echo arrival time and the propagation time reference plane is calculated to generate a propagation time difference sequence, identify local propagation time advance segments, and determine the propagation time advance amount; Based on cross-frequency homo-source reflection trajectories, local propagation time advance segments with corresponding positions and consistent propagation time advance trends under different detection frequency bands are connected to generate cross-frequency propagation time difference vacancy chains and vacancy chain spatial characterization data. Based on the spatial characterization data of the vacancy chain, the equivalent thickness of the vacancy is generated, and the equivalent thickness of the vacancy is mapped to the underground space. The mapping results of different underground media interfaces and adjacent continuous detection paths are superimposed and the boundaries are closed to generate the propagation time difference vacancy closed body. Based on the echo response correction propagation time difference vacancy closed body corresponding to the top and bottom boundaries, the underground cavity region is generated, and the location, burial depth and spatial range of the underground cavity are determined.

[0006] Optionally, the generation of the multi-frequency underground detection image set includes: Multiple continuous detection paths are laid out along the underground area to be detected, and detection positions are determined on each continuous detection path; Multiple detection frequency bands were used to collect underground echo data at each detection location, and the propagation velocity in normal medium, the propagation velocity in cavitary medium, and the propagation direction of the detection signal were obtained. The underground echo data is aligned in position, corrected at zero time, and corrected in frequency band response to generate corrected underground echo data. The corrected underground echo data is then arranged according to the detection location and detection frequency band to generate a multi-frequency underground detection image group.

[0007] Optionally, the generation of the cross-frequency co-source reflection trajectory and trajectory characterization data includes: By traversing the multi-frequency underground detection image group according to the detection frequency band, and connecting the reflection pixels with continuous depth changes and connected echo phases in adjacent detection locations, a continuous reflection ridge is generated. The continuous reflection ridges are correlated with the detection location, detection frequency band, echo arrival time, and echo phase to generate reflection characterization data. Based on the reflection characterization data, the spatial positions of continuous reflection ridges corresponding to different detection frequency bands are mapped, and spatial position mapping results are generated. Compare the depth arrangement order and echo phase change of the continuous reflection ridges corresponding to different detection frequency bands to generate depth arrangement correspondence results and echo response correspondence results; By connecting continuous reflection ridges with consistent spatial location, depth arrangement, and echo response, cross-frequency homogeneous reflection trajectories are generated, and the corresponding reflection characterization data are collected to generate trajectory characterization data.

[0008] Optionally, the generation of the propagation time reference surface includes: The detection location, detection frequency band, and echo arrival time are obtained from the trajectory characterization data, and the trajectory time point set is generated by arranging them according to the detection location and detection frequency band. Using the detection location, detection frequency band, and echo arrival time as the location coordinates, frequency band coordinates, and time coordinates, respectively, an initial surface is fitted to the trajectory time point set to generate an initial propagation time surface; Calculate the surface fitting residual between each trajectory time point and the initial propagation time surface, and compare the fitting residuals of adjacent surfaces along the detection position and detection frequency band to generate position continuity data and frequency band consistency data; By combining location continuity data and frequency band consistency data to determine fitting weights, a weighted surface is fitted to the trajectory time point set, and the surface fitting residuals and fitting weights are updated until the time value changes of two adjacent weighted propagation time surfaces remain stable, thus generating a propagation time trend surface. By associating the propagation time trend surface with the corresponding cross-frequency homologous reflection trajectory, a propagation time reference surface corresponding to the underground medium interface is generated.

[0009] Optionally, determining the propagation time advance includes: Based on the detection location and frequency band, the actual echo arrival time is mapped to the propagation time reference plane to determine the corresponding reference echo arrival time, and the difference between the reference echo arrival time and the actual echo arrival time is calculated to generate the propagation time difference value. The propagation time difference values ​​are associated with the detection location and detection frequency band, and arranged along the detection location according to the detection frequency band to generate a propagation time difference sequence; Calculate the time difference change between adjacent detection positions in the propagation time difference sequence, and arrange them along the detection positions to generate a propagation time difference change sequence; By combining the propagation time difference sequence and the propagation time difference change sequence, segments with continuously positive and continuously changing propagation time difference values ​​are identified, isolated positive propagation time difference values ​​appearing at a single detection location are eliminated, the segment start point and segment end point are determined, and local propagation time advance segments are generated. The propagation time difference value corresponding to each detection position in the local propagation time advance segment is determined as the propagation time advance amount, and associated with the corresponding local propagation time advance segment.

[0010] Optionally, the generation of the vacancy chain spatial representation data includes: Taking each cross-frequency co-source reflection trajectory as the processing object, local propagation time advance segments are arranged according to the detection frequency band, and the corresponding detection positions and propagation time advance are associated to generate a cross-frequency advance segment set; The local propagation time advance segments corresponding to adjacent detection frequency bands are matched with the same detection position to determine the overlapping range of detection positions between segments and generate horizontal position correspondence data. Arrange the propagation time advance within the overlapping interval of the detection positions according to the detection positions, compare the direction of increase and decrease and the turning point of the propagation time advance in adjacent detection frequency bands, and generate corresponding data on the propagation time advance trend. By combining the data corresponding to the horizontal position and the data corresponding to the propagation time lead trend, the local propagation time lead segments that correspond to the horizontal position and have the same trend of propagation time lead change are connected sequentially to generate the cross-frequency propagation time difference gap chain. Based on the cross-frequency propagation time difference vacancy chain, the lateral extension range and propagation time advance distribution are determined. The depth of the subsurface medium interface is generated by combining the echo arrival time and normal medium propagation velocity in the corresponding cross-frequency co-source reflection trajectory, and then correlated to generate vacancy chain spatial characterization data.

[0011] Optionally, the generation of the propagation time difference vacancy closure includes: Arrange the propagation time advance corresponding to each detection position according to the propagation time advance distribution, calculate the time difference per unit propagation distance corresponding to the propagation speed of normal medium and the propagation speed of void medium, and convert the propagation time advance corresponding to each detection position to generate the void equivalent thickness. The reverse mapping starting point corresponding to each detection location is determined based on the depth of the underground medium interface. The equivalent thickness of the vacancy is extended in the opposite direction of the detection signal propagation to determine the reverse mapping endpoint. Within the lateral extension range, the inverse mapping start point and inverse mapping end point corresponding to adjacent detection locations are connected respectively to generate the inverse mapping start boundary and inverse mapping end boundary. The underground space between the two boundaries is then processed to generate a propagation time difference space candidate projection body, which is then associated with the corresponding underground medium interface and continuous detection path. Candidate projections of propagation time difference vacancy corresponding to different underground media interfaces and adjacent continuous detection paths are placed in the same underground space, and overlapping candidate projections of propagation time difference vacancy are connected to generate a candidate projection superposition area. Identify the outer edge of the candidate projection superposition area, generate the outer boundary, connect the continuously distributed outer boundaries between different underground media interfaces and between adjacent continuous detection paths, and close the discontinuous positions to generate a propagation time difference vacancy closure body.

[0012] Optionally, the generation of the location, depth, and spatial extent of the underground cavity includes: The echo position is determined according to the top and bottom boundaries of the propagation time difference space closed body, the corresponding echo amplitude and echo phase are obtained, and their changes are compared along the detection position to generate the top interface response and bottom interface response. Based on the top and bottom interface responses, the top and bottom interface locations where the echo amplitude changes abruptly and the echo phase changes are identified, respectively. The top and bottom boundaries of the propagation time difference void closure body are corrected by using the top and bottom interface positions respectively, while the lateral boundaries are preserved, to generate underground void regions; Based on the top, bottom, and lateral boundaries of the underground cavity area, the corresponding depth, detection location range, and continuous detection path are determined to generate the location, burial depth, and spatial range of the underground cavity.

[0013] An automatic underground cavity detection system based on image recognition according to an embodiment of the present invention includes: The image generation module is used to acquire and correct underground echo data and generate multi-frequency underground detection image sets; The trajectory generation module is used to identify and connect continuous reflection ridges corresponding to different detection frequency bands, and generate cross-frequency homogeneous reflection trajectories and trajectory characterization data; The reference surface generation module is used to perform surface fitting on the changing trend of echo arrival time and generate a propagation time reference surface. The advance segment generation module is used to generate local propagation time advance segments and propagation time advance amounts; The vacancy chain generation module is used to connect local propagation time advance segments to generate cross-frequency propagation time difference vacancy chains and vacancy chain spatial representation data; The closed volume generation module is used to generate the equivalent thickness of the void and the candidate projection of the propagation time difference void, and to perform spatial superposition and boundary closure to generate the propagation time difference void closed volume; The cavity detection module is used to correct the top and bottom boundaries of the propagation time difference vacancy closure body, and generate underground cavity areas as well as the location, burial depth and spatial range of underground cavities.

[0014] The beneficial effects of this invention are: This invention acquires underground echo data using multiple detection frequency bands along multiple continuous detection paths, and performs position alignment, time zero-point correction, and frequency band response correction on the underground echo data. This reduces data offset between different detection locations and frequency bands, providing a unified data foundation for multi-frequency underground detection image groups. By identifying continuous reflection ridges and combining spatial location, depth arrangement, and echo response to correspond and connect continuous reflection ridges under different detection frequency bands, a cross-frequency homogeneous reflection trajectory reflecting the cross-frequency change process of the same underground medium interface can be formed. This reduces the interference of clutter and local anomaly responses in a single detection frequency band on the underground cavity identification results.

[0015] This invention constructs a propagation time reference surface based on the echo arrival time in cross-frequency co-source reflection trajectories, and compares the actual echo arrival time with the propagation time reference surface. This allows for the separation of local propagation time advance phenomena from the overall change trend of the underground medium interface. By identifying local propagation time advance segments and connecting these segments with corresponding lateral positions and consistent propagation time advance trends across different detection frequency bands into cross-frequency propagation time difference vacancy chains, propagation anomalies dispersed across different detection frequency bands can be continuously correlated. This reduces false identification caused by isolated single-frequency anomalies and improves the reliability of determining underground cavity anomaly areas.

[0016] This invention further combines the propagation velocity of normal media and the propagation velocity of cavitary media, converting the propagation time advance into the equivalent thickness of the vacancy, and mapping it to the underground space in the opposite direction of the probe signal propagation, generating a candidate projection of the propagation time difference vacancy. By spatially superimposing and boundary-closing the candidate projections of the propagation time difference vacancy corresponding to different underground media interfaces and adjacent probe paths, the overlapping relationship of multi-interface and multi-path propagation anomalies in the underground space can be used to determine the closed body of the propagation time difference vacancy, realizing the conversion from propagation time anomalies to the range of underground cavity space, overcoming the problem that it is difficult to determine the complete boundary by relying solely on direct reflection from the cavity.

[0017] This invention also utilizes the echo amplitude and echo phase corresponding to the top and bottom boundaries of the propagation time difference void closure body to generate top and bottom interface responses, and corrects the top and bottom boundaries of the propagation time difference void closure body accordingly, allowing the propagation time inversion results to be cross-validated with the actual interface echo responses. This further enables the determination of the location, depth, and spatial extent of underground cavities, improving the completeness of underground cavity boundary determination and the accuracy of detection results, and providing reliable data support for cavity risk assessment in underground roads, building foundations, and underground facility areas. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of an automatic detection method for underground cavities based on image recognition proposed in this invention; Figure 2 This is a schematic diagram of the cross-frequency propagation time difference void chain and the spatial representation data of the void chain in the automatic detection method of underground cavities based on image recognition proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 An automatic detection method for underground cavities based on image recognition includes the following steps: Multiple continuous detection paths are deployed along the underground area to be detected. Detection positions are determined on each continuous detection path. Underground echo data containing echo arrival time, echo amplitude, and echo phase are collected using multiple detection frequency bands. The propagation velocity of normal medium, propagation velocity of cavity medium, and propagation direction of detection signal are obtained. The underground echo data is then aligned in position, corrected for time zero point, and corrected for frequency band response to generate a multi-frequency underground detection image group. Identify continuous reflection ridges in multi-frequency underground detection image groups, associate continuous reflection ridges with corresponding detection locations, detection frequency bands, echo arrival times, and echo phases to generate reflection characterization data. Based on the reflection characterization data corresponding to different detection frequency bands, perform spatial location correspondence, depth arrangement correspondence, and echo response correspondence on continuous reflection ridges. Connect continuous reflection ridges with consistent correspondence to generate cross-frequency co-source reflection trajectories and trajectory characterization data. Based on trajectory characterization data, surface fitting is performed on the overall variation trend of echo arrival time of each cross-frequency co-source reflection trajectory along the detection position and detection frequency band to generate the propagation time reference surface corresponding to each underground medium interface. The difference between the actual echo arrival time and the propagation time reference plane in the cross-frequency co-source reflection trajectory is calculated to generate a propagation time difference sequence. The segments in the propagation time difference sequence where the echo arrival time is continuously advanced are identified, local propagation time advance segments are generated, and the propagation time advance amount is determined. Based on the cross-frequency homogeneous reflection trajectory, local propagation time advance segments with corresponding lateral positions and consistent propagation time advance trends under different detection frequency bands are connected to generate cross-frequency propagation time difference space chains. The cross-frequency propagation time difference space chains are then spatially characterized to generate space chain spatial characterization data. Based on the spatial characterization data of the vacancy chain, the propagation velocity of the normal medium and the propagation velocity of the cavity medium, the equivalent thickness of the vacancy is generated. Combined with the propagation direction of the detection signal, the equivalent thickness of the vacancy is back-mapped to the underground space to generate a candidate projection body of the propagation time difference vacancy. The candidate projection bodies of the propagation time difference vacancy corresponding to different underground medium interfaces and adjacent detection paths are spatially superimposed and boundary closed to generate a closed body of the propagation time difference vacancy. Identify the echo responses corresponding to the top and bottom boundaries of the propagation time difference void closure body, generate the top and bottom interface responses, correct the top and bottom boundaries of the propagation time difference void closure body based on the top and bottom interface responses, generate the underground cavity region, determine the depth corresponding to the top and bottom boundaries of the underground cavity region and the detection location range and detection path corresponding to the lateral boundaries, and generate the location, burial depth and spatial range of the underground cavity.

[0021] In this embodiment, the generation of multi-frequency underground detection image sets includes: Multiple continuous detection paths are laid out along the underground area to be detected, and detection positions are determined on each continuous detection path; Multiple detection frequency bands were used to collect underground echo data at each detection location, and the propagation velocity in normal medium, the propagation velocity in cavitary medium, and the propagation direction of the detection signal were obtained. The underground echo data is aligned in position, corrected at zero time, and corrected in frequency band response to generate corrected underground echo data. The corrected underground echo data is then arranged according to the detection location and detection frequency band to generate a multi-frequency underground detection image group.

[0022] In this embodiment, the generation of cross-frequency co-source reflection trajectories and trajectory characterization data includes: According to the detection frequency band, the multi-frequency underground detection image group is traversed, and the reflection pixels with continuous depth changes and connected echo phases in adjacent detection positions are identified. The reflection pixels are then connected along the detection positions to generate continuous reflection ridges corresponding to each detection frequency band. Determine the detection position, detection frequency band, echo arrival time and echo phase of each reflection pixel in the continuous reflection ridge, arrange each reflection pixel and its corresponding data according to the detection position, and generate reflection characterization data corresponding to each continuous reflection ridge. Based on the reflection characterization data, the continuous reflection ridges corresponding to different detection frequency bands are matched according to the same detection position. The spatial position differences of the continuous reflection ridges at the corresponding detection positions are compared to generate spatial position correspondence results. Specifically, the continuous reflection ridges corresponding to 200MHz, 400MHz, and 800MHz in the same continuous detection path are mapped to the same coordinate plane composed of the detection position coordinates and the echo arrival time coordinates. 201 detection positions are formed at 0.05m intervals. The longitudinal image coordinate difference of different continuous reflection ridges at the same detection position is calculated. The longitudinal image coordinate difference corresponds to the echo arrival time difference. The average echo arrival time difference of the continuous reflection ridges corresponding to 200MHz and 400MHz is 0.60ns, and the average echo arrival time difference of the continuous reflection ridges corresponding to 400MHz and 800MHz is 0.68ns. The continuous reflection ridges are matched according to the cumulative echo arrival time difference in ascending order to generate spatial position correspondence results. According to the echo arrival time, the continuous reflection ridges in the same detection frequency band are arranged in depth. The depth arrangement order of the continuous reflection ridges along the detection position in different detection frequency bands is compared to generate the corresponding depth arrangement results. The echo phase polarity, phase inflection position and phase change direction of the corresponding continuous reflection ridges are compared to generate the corresponding echo response results. Specifically, continuous reflection ridges are arranged at each detection location in ascending order of echo arrival time to form a depth arrangement order corresponding to each detection frequency band. Among the 201 detection locations, 188 detection locations have the same depth arrangement order. The echo phase polarity, phase inflection position, and phase change direction are compared sequentially for the continuous reflection ridges corresponding to the spatial locations. The average difference of the phase inflection positions is 0.05m. There are 186 detection locations with the same echo phase change direction. The depth arrangement corresponding results and echo response corresponding results are generated respectively. Continuous reflection ridges with consistent spatial location, depth arrangement, and echo response results are connected according to the detection frequency band to generate cross-frequency homogeneous reflection trajectories. Reflection characterization data corresponding to the continuous reflection ridges connected by the cross-frequency homogeneous reflection trajectories are collected, and the reflection characterization data are arranged according to the detection location and detection frequency band to generate trajectory characterization data. Specifically, the spatial location correspondence results, depth arrangement correspondence results, and echo response correspondence results are associated according to the detection location to determine the detection locations where all three correspondence results are consistent. Among the 201 detection locations, 117 detection locations have consistent results across all three categories. These 117 detection locations continuously cover the detection location range from 4.10m to 9.90m. The continuous reflection ridges corresponding to the 117 detection locations are connected in ascending order of detection frequency bands to generate cross-frequency homogeneous reflection trajectories. Each detection location corresponds to reflection characterization data for three detection frequency bands, forming a total of 351 sets of cross-frequency echo data. The 351 sets of cross-frequency echo data are arranged according to the detection location and detection frequency band to generate trajectory characterization data.

[0023] In this embodiment, the generation of the propagation time reference plane includes: The detection location, detection frequency band, and echo arrival time are obtained from the trajectory characterization data corresponding to each cross-frequency homogeneous reflection trajectory. The echo arrival times are arranged according to the detection location and detection frequency band to generate the trajectory time point set corresponding to each cross-frequency homogeneous reflection trajectory. Using the detection location in the trajectory time point set as the location coordinate, the detection frequency band as the frequency band coordinate, and the echo arrival time as the time coordinate, the trajectory time point set is fitted with an initial surface to generate an initial propagation time surface. Specifically, a surface fitting was performed on 117 detection locations and 351 trajectory time points corresponding to the three detection frequency bands covered by the cross-frequency co-source reflection trajectory. A surface expression containing a first-order term for the detection location, a second-order term for the detection location, a first-order term for the detection frequency band, a second-order term for the detection frequency band, and a product term of the detection location and the detection frequency band was adopted. The coefficients of each term in the surface expression were adjusted to minimize the sum of squares of the differences between the 351 trajectory time points and the corresponding time values ​​on the surface. The average time difference obtained by fitting was 0.47 ns, and the initial propagation time surface was generated. Calculate the difference between the echo arrival time and the corresponding time value of the initial propagation time curve at each trajectory time point, generate the surface fitting residual, compare the surface fitting residuals of adjacent trajectory time points in the same detection frequency band along the detection position, and compare the surface fitting residuals of adjacent trajectory time points in the same detection position along the detection frequency band to generate position continuity data and frequency band consistency data. Specifically, the difference between the echo arrival time and the corresponding time value of the initial propagation time surface at 351 trajectory time points is calculated to generate surface fitting residuals. The difference between the surface fitting residuals corresponding to adjacent trajectory time points in the same detection frequency band is calculated along the detection position, and the average difference is 0.12 ns, generating position continuity data. The difference between the surface fitting residuals corresponding to adjacent trajectory time points in the same detection position is calculated along the detection frequency band, and the average difference is 0.18 ns, generating frequency band consistency data. By combining location continuity data and frequency band consistency data, the fitting weights for each trajectory time point are determined. Based on the fitting weights, a weighted surface is fitted to the trajectory time point set to generate a weighted propagation time surface. The difference between the echo arrival time of each trajectory time point and the corresponding time value of the weighted propagation time surface is calculated, and the surface fitting residual is updated. The fitting weights are re-determined using the updated surface fitting residual, and the weighted propagation time surface is regenerated until the time value changes of the weighted propagation time surface generated in two adjacent times remain stable. The weighted propagation time surface generated last is determined as the propagation time trend surface. Specifically, the surface fitting residuals, position continuity data, and frequency band consistency data are normalized respectively. The sum of the three normalization results is added by one and the reciprocal is taken to generate the fitting weights for each trajectory time point. Based on the fitting weights, a weighted surface fitting is performed on 351 trajectory time points to generate a weighted propagation time surface. The surface fitting residuals are recalculated, and the position continuity data and frequency band consistency data are recalculated based on the updated surface fitting residuals. Then, the fitting weights are regenerated. Six consecutive weighted surface fittings are performed. The maximum time value changes between the initial propagation time surface and the first weighted propagation time surface, as well as between subsequent adjacent weighted propagation time surfaces, are 0.62ns, 0.31ns, 0.17ns, 0.08ns, 0.05ns, and 0.03ns, respectively. After retaining each time value to 0.1ns, the corresponding time values ​​of the weighted propagation time surfaces generated in the fifth and sixth times are consistent. The weighted propagation time surface generated in the sixth time is determined as the propagation time trend surface. By associating the propagation time trend surface with the corresponding cross-frequency homo-source reflection trajectory, a propagation time reference surface corresponding to the subsurface medium interface represented by each cross-frequency homo-source reflection trajectory is generated. Specifically, the propagation time trend surface is associated with the corresponding cross-frequency co-source reflection trajectory. The corresponding underground medium interface is determined according to the arrangement of the echo arrival times of the cross-frequency co-source reflection trajectory. The average echo arrival times of the three cross-frequency co-source reflection trajectories are 18.4ns, 27.7ns, and 39.2ns, respectively, corresponding to the three underground medium interfaces. The three propagation time trend surfaces are associated with the corresponding underground medium interfaces to generate the propagation time reference surfaces corresponding to each underground medium interface.

[0024] In this embodiment, determining the propagation time advance includes: According to the detection location and detection frequency band, the actual echo arrival time in the cross-frequency co-source reflection trajectory is mapped to the propagation time reference surface of the corresponding underground medium interface, the reference echo arrival time corresponding to the same detection location and detection frequency band is determined, and the propagation time difference value obtained by subtracting the actual echo arrival time from the reference echo arrival time is calculated. The propagation time difference value is associated with the corresponding detection position and detection frequency band, and the propagation time difference value is arranged along the detection position according to the detection frequency band to generate a propagation time difference sequence; Calculate the time difference change between adjacent detection positions in the propagation time difference sequence, arrange the time difference changes along the detection positions, and generate a propagation time difference change sequence; By combining the propagation time difference sequence and the propagation time difference change sequence, segments with continuously positive propagation time difference values ​​that change continuously along the detection position are identified. Segments with positive propagation time difference values ​​only appearing at a single detection position and non-positive propagation time difference values ​​at adjacent detection positions are eliminated. The detection position where the propagation time difference value changes from non-positive to positive is taken as the starting point of the segment, and the detection position before the propagation time difference value changes from positive to non-positive is taken as the ending point of the segment, thus generating a local propagation time advance segment. Specifically, propagation time difference sequences corresponding to 200MHz, 400MHz, and 800MHz were read along the detection positions. The detection positions where the propagation time difference value changed from negative to positive were recorded as the starting positions of the continuous segments, and the detection positions before the propagation time difference value changed from positive to negative were recorded as the ending positions of the continuous segments. The propagation time difference sequence corresponding to 200MHz formed a single detection position segment at 1.60m, 24 detection position segments covering 4.20m to 5.35m, and 16 detection positions covering 7.10m to 7.85m. The propagation time difference sequence corresponding to 400MHz forms a single detection position segment at 2.35m and 19 detection position segments covering 4.30m to 5.20m. The propagation time difference sequence corresponding to 800MHz forms 17 detection position segments covering 4.35m to 5.15m and a single detection position segment at 8.40m. After removing three segments containing only a single detection position, four local propagation time advance segments with lengths of 1.15m, 0.75m, 0.90m and 0.80m are generated. The propagation time difference value corresponding to each detection position in the local propagation time advance segment is determined as the propagation time advance, and the propagation time advance is associated with the corresponding local propagation time advance segment according to the detection position.

[0025] In this embodiment, the generation of vacancy chain spatial representation data includes: Taking each cross-frequency co-source reflection trajectory as the processing object, the corresponding local propagation time advance segments are arranged according to the detection frequency band, and the local propagation time advance segments are associated with the corresponding detection positions and propagation time advance amounts to generate a cross-frequency advance segment set; The local propagation time advance segments corresponding to adjacent detection frequency bands in the cross-frequency advance segment set are matched according to the same detection position. The detection positions of the segment start point, segment end point, and the detection positions between the segment start point and segment end point are compared to determine the detection position overlap interval between local propagation time advance segments. The detection position overlap interval is associated with the corresponding local propagation time advance segment to generate horizontal position correspondence data. Specifically, the local propagation time advance segments corresponding to adjacent detection frequency bands are mapped to the same detection position axis. The local propagation time advance segment corresponding to 200MHz covers 4.20m to 5.35m, and the local propagation time advance segment corresponding to 400MHz covers 4.30m to 5.20m. The two form a detection position overlap interval of 4.30m to 5.20m. The local propagation time advance segment corresponding to 400MHz and the local propagation time advance segment corresponding to 800MHz form a detection position overlap interval of 4.35m to 5.15m. The detection position overlap interval is associated with the corresponding local propagation time advance segment to generate horizontal position corresponding data. Obtain the overlapping interval of detection positions from the data corresponding to the horizontal positions, arrange the propagation time advance within the overlapping interval of detection positions according to the detection positions, compare the direction of increase and decrease and the turning point of the propagation time advance in adjacent detection frequency bands, and generate the corresponding data of propagation time advance trend. Specifically, the propagation time advance within the overlapping interval of the detection positions is arranged according to the same detection position. The overlapping interval of the detection positions corresponding to 200MHz and 400MHz contains 19 detection positions, and the difference between 18 adjacent propagation time advances is calculated. Among them, the positive and negative directions of 17 of the difference are consistent, and the turning points of the two increasing and decreasing direction sequences differ by 0.05m. The overlapping interval of the detection positions corresponding to 400MHz and 800MHz contains 17 detection positions, and the difference between 16 of the difference is calculated. Among them, the positive and negative directions of 15 of the difference are consistent, and the turning points of the two increasing and decreasing direction sequences differ by 0.10m. The corresponding increasing and decreasing directions and turning points are correlated to generate propagation time advance trend data. By combining the data corresponding to the horizontal position and the data corresponding to the propagation time advance trend, the local propagation time advance segments that correspond to the horizontal position and have the same propagation time advance change trend are connected sequentially according to the order of the detection frequency bands, and the local propagation time advance segments connected between adjacent detection frequency bands are continuously connected to generate a cross-frequency propagation time difference gap chain. Specifically, using a 200MHz local propagation time advance segment covering 4.20m to 5.35m as the connection starting point, within the overlapping detection positions from 4.30m to 5.20m, data from the corresponding 200MHz and 400MHz segments at 19 identical detection positions are connected to form 19 sets of cross-frequency segment connection data. Within the overlapping detection positions from 4.35m to 5.15m, data from the corresponding 400MHz and 800MHz segments at 17 identical detection positions are connected to form 17 sets of cross-frequency segment connection data. Each set of cross-frequency segment connection data is connected sequentially along the detection positions, and the two connection results are connected through the local propagation time advance segment corresponding to 400MHz to generate a cross-frequency propagation time difference gap chain. The lateral extension range is determined based on the local propagation time advance segments connected by the cross-frequency propagation time difference space chain. The propagation time advance is collected according to the detection location to generate the propagation time advance distribution. The depth of the subsurface medium interface is generated based on the echo arrival time and normal medium propagation velocity in the corresponding cross-frequency co-source reflection trajectory. The lateral extension range, the propagation time advance distribution and the subsurface medium interface depth are correlated to generate space chain spatial characterization data. Specifically, local propagation time advance segments in the cross-frequency propagation time difference vacancy chain are mapped to the detection position axis. A lateral extension range of 4.20m to 5.35m is determined based on the starting point of the first segment and the ending point of the last segment. Within this lateral extension range, the propagation time advance associated with the cross-frequency propagation time difference vacancy chain at each detection position is collected. The average value of the propagation time advance collected at the same detection position is calculated. For detection positions associated with only one detection frequency band, the propagation time advance corresponding to that detection frequency band is used to generate a propagation time advance distribution that increases from 0.42ns at the starting position to 1.86ns in the middle and then decreases to 0.38ns at the ending position. The echo arrival time of 27.7ns is halved and multiplied by the normal medium propagation velocity of 0.12m / ns to generate a subsurface medium interface depth of 1.662m. The lateral extension range, the propagation time advance distribution, and the subsurface medium interface depth are correlated to generate vacancy chain spatial characterization data.

[0026] In this embodiment, the generation of the propagation time difference vacancy closure body includes: Arrange the propagation time advance corresponding to each detection position according to the propagation time advance distribution in the spatial characterization data of the vacancy chain, calculate the unit propagation distance time difference corresponding to the propagation speed of the normal medium and the propagation speed of the vacancy medium, generate the unit propagation distance time difference, use the unit propagation distance time difference to convert the propagation time advance corresponding to each detection position, and generate the vacancy equivalent thickness corresponding to each detection position. Specifically, the propagation speed in normal medium is 0.12 m / ns, corresponding to a round-trip time of 16.67 ns / m per unit propagation distance. The propagation speed in void medium is 0.30 m / ns, corresponding to a round-trip time of 6.67 ns / m per unit propagation distance. The difference between the two is calculated to generate a time difference of 10.00 ns / m per unit propagation distance. The propagation time advance of each detection position is divided by the time difference per unit propagation distance. A position with a propagation time advance of 1.86 ns generates a vacancy equivalent thickness of 0.186 m, and a position with a propagation time advance of 0.42 ns generates a vacancy equivalent thickness of 0.042 m. The vacancy equivalent thicknesses are arranged according to the detection positions. The reverse mapping starting point corresponding to each detection location is determined according to the subsurface medium interface depth in the vacancy chain spatial characterization data, and the vacancy equivalent thickness is extended in the opposite direction of the detection signal propagation to determine the reverse mapping endpoint. Specifically, the reverse mapping starting point is determined according to the detection location and the underground medium interface depth of 1.662m, and each reverse mapping starting point is associated with the corresponding vacancy equivalent thickness. When the detection signal propagates vertically downward, the reverse mapping starting point moves vertically upward along the corresponding vacancy equivalent thickness. The position with a vacancy equivalent thickness of 0.186m forms a reverse mapping endpoint with a depth of 1.476m, and the position with a vacancy equivalent thickness of 0.042m forms a reverse mapping endpoint with a depth of 1.620m. Within the lateral extension range of the spatial representation data of the vacancy chain, the inverse mapping starting point corresponding to the adjacent detection location is connected to generate the inverse mapping starting boundary, and the inverse mapping ending point corresponding to the adjacent detection location is connected to generate the inverse mapping ending boundary. The underground space between the inverse mapping starting boundary and the inverse mapping ending boundary is processed to generate the propagation time difference vacancy candidate projection, and the propagation time difference vacancy candidate projection is associated with the corresponding underground medium interface and continuous detection path. Specifically, reverse mapping is performed based on the spatial characterization data of the vacancy chains corresponding to each continuous detection path and each underground medium interface. The lateral extension range of the third continuous detection path and the second underground medium interface is 4.20m to 5.35m. The 24 reverse mapping starting points within this lateral extension range are connected in the order of detection positions to generate the reverse mapping starting boundary. The 24 reverse mapping ending points are connected in the same order to generate the reverse mapping ending boundary. The reverse mapping starting points and ending points corresponding to adjacent detection positions are connected to form 23 continuous projections. The underground space surrounded by the reverse mapping starting boundary, the reverse mapping ending boundary and the 23 continuous projections is processed to generate the propagation time difference vacancy candidate projections corresponding to the third continuous detection path and the second underground medium interface. The same process is used to generate the propagation time difference vacancy candidate projections corresponding to the other continuous detection paths and underground medium interfaces. Based on the underground medium interface and continuous detection path corresponding to the propagation time difference spatial candidate projection, the propagation time difference spatial candidate projections corresponding to different underground medium interfaces and adjacent continuous detection paths are placed in the same underground space, and the overlapping propagation time difference spatial candidate projections are connected to generate a candidate projection superposition area. Specifically, six continuous detection paths are numbered sequentially along the direction of their arrangement, with a spacing of 0.10m between adjacent continuous detection paths. The starting detection position of the first continuous detection path is taken as the starting position on the plane. The extension direction and the arrangement direction of the continuous detection paths are taken as two horizontal coordinate directions, and the underground depth direction is taken as the vertical coordinate direction. According to the corresponding continuous detection path, detection position, and underground medium interface depth, 18 candidate projections of propagation time difference vacancy corresponding to the six continuous detection paths and three underground medium interfaces are placed in the same underground space. The spatial overlap between the candidate projections of propagation time difference vacancy corresponding to different underground medium interfaces and the candidate projections of propagation time difference vacancy corresponding to adjacent continuous detection paths is calculated. The 14 spatial overlaps with common space or common boundary are connected sequentially to generate a candidate projection superposition area. Identify the outer edge of the candidate projection superposition area, generate the outer boundary, connect the continuously distributed outer boundaries between different underground media interfaces and between adjacent continuous detection paths, and close the discontinuities in the outer boundary to generate a propagation time difference vacancy closure body. Specifically, the outer edge of the candidate projection superposition area is identified, and an outer boundary consisting of 196 continuous outer surfaces is generated. For the nine discontinuous positions in the outer boundary, the connection continues along the extension direction of the outer surfaces on both sides of the discontinuity position until the connection results on both sides intersect, closing the discontinuous positions in the outer boundary. The underground space enclosed by the closed outer boundary is then filled to generate a propagation time difference space closure body with a length of 1.08m, a width of 0.46m, and a maximum height of 0.186m.

[0027] In this embodiment, the generation of the location, depth, and spatial extent of the underground cavity includes: The corresponding echo positions are determined according to the top and bottom boundaries of the propagation time difference space closure body. The echo amplitude and echo phase corresponding to each echo position are obtained from the underground echo data. The changes in echo amplitude and echo phase are compared according to the order of the detection positions to generate the top interface response and bottom interface response. Specifically, the top and bottom boundaries of the propagation time difference spatial closure body are mapped to six continuous detection paths. On each continuous detection path, the detection positions covered by the top and bottom boundaries are determined at 0.05m intervals, resulting in a total of 144 top boundary echo positions and 144 bottom boundary echo positions. The echo amplitude and echo phase are obtained from three sampling points before and after the echo arrival time corresponding to each echo position, as well as the current sampling point. The top and bottom boundaries generate 1008 echo amplitude data and 1008 echo phase data, respectively. The data corresponding to the top boundary are arranged according to the continuous detection paths and detection positions to generate the top interface response. The data corresponding to the bottom boundary are arranged in the same way to generate the bottom interface response. The top interface location is identified based on the top interface response, where the echo amplitude changes continuously and the echo phase changes. The bottom interface location is identified based on the bottom interface response, where the echo amplitude changes continuously and the echo phase changes. Specifically, the echo positions where the echo phase changes are determined in the top interface responses corresponding to the 144 top boundary echo positions, and the sum of the echo amplitude changes of each echo position and the sampling points before and after it is calculated. The echo position with the largest sum of echo amplitude changes in each top boundary echo position is determined as the corresponding top interface position, generating 144 top interface positions. The same processing procedure is used to process the bottom interface responses corresponding to the 144 bottom boundary echo positions, generating 144 bottom interface positions. The average depth of the 144 top interface positions is 1.470m, and the average depth of the 144 bottom interface positions is 1.646m. The top boundary of the propagation time difference void closure body is corrected by using the top interface position, the bottom boundary of the propagation time difference void closure body is corrected by using the bottom interface position, and the lateral boundary of the propagation time difference void closure body is preserved to generate an underground void region. Specifically, according to the detection location, the top interface positions corresponding to each continuous detection path are connected to generate six top interface connection lines. The top interface positions corresponding to the detection locations in adjacent continuous detection paths are then connected to generate the corrected top boundary. According to the detection location, the bottom interface positions corresponding to each continuous detection path are connected to generate six bottom interface connection lines. The bottom interface positions corresponding to the detection locations in adjacent continuous detection paths are then connected to generate the corrected bottom boundary. The average depths of the top and bottom boundaries before correction are 1.482m and 1.660m, respectively. The average depths of the top and bottom boundaries after correction are 1.470m and 1.646m, respectively. The corrected top boundary, the corrected bottom boundary, and the lateral boundary of the propagation time difference void closure body are then connected to generate the underground cavity region. Based on the underground cavity area, the depths corresponding to the top and bottom boundaries of the underground cavity area are determined, and the detection location range and continuous detection path corresponding to the lateral boundaries of the underground cavity area are determined, thereby generating the location, burial depth and spatial range of the underground cavity. Specifically, the center of the underground cavity area in the direction of the ground surface is aligned with the 4.78m detection position of the third continuous detection path to generate the location of the underground cavity. The average depth corresponding to 144 top interface positions is calculated to generate an underground cavity burial depth of 1.470m. The maximum span of the underground cavity area along the direction of the continuous detection path, the direction of the continuous detection path arrangement, and the direction of underground depth is calculated respectively to generate a spatial range with a length of 1.08m, a width of 0.46m, and a height of 0.18m.

[0028] An automatic underground cavity detection system based on image recognition includes: The image generation module is used to acquire and correct underground echo data and generate multi-frequency underground detection image sets; The trajectory generation module is used to identify and connect continuous reflection ridges corresponding to different detection frequency bands, and generate cross-frequency homogeneous reflection trajectories and trajectory characterization data; The reference surface generation module is used to perform surface fitting on the changing trend of echo arrival time and generate a propagation time reference surface. The advance segment generation module is used to generate local propagation time advance segments and propagation time advance amounts; The vacancy chain generation module is used to connect local propagation time advance segments to generate cross-frequency propagation time difference vacancy chains and vacancy chain spatial representation data; The closed volume generation module is used to generate the equivalent thickness of the void and the candidate projection of the propagation time difference void, and to perform spatial superposition and boundary closure to generate the propagation time difference void closed volume; The cavity detection module is used to correct the top and bottom boundaries of the propagation time difference vacancy closure body, and generate underground cavity areas as well as the location, burial depth and spatial range of underground cavities.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the investigation of underground cavities in a subsidence area of ​​an in-service road. This area has long been affected by vehicle loads and rainwater infiltration, resulting in slight road surface subsidence, cracks, and localized water accumulation. Maintenance personnel suspected underground cavities formed by the loss of backfill material beneath the road structure. Since the road still needed to be passable, large-scale excavation was not feasible. Furthermore, underground pipelines, water-bearing areas, loose backfill areas, and undulating media interfaces could all generate abnormal echoes. Using a single detection frequency band could easily lead to misjudgments and make it difficult to fully determine the top, bottom, and lateral boundaries of the underground cavities.

[0030] Six parallel continuous detection paths were deployed along the area to be detected, with a distance of 0.10m between adjacent paths. Each path was 10m long, and 201 detection positions were established at 0.05m intervals, resulting in a total of 1206 detection positions. Subsurface echo data were collected using three detection frequency bands: 200MHz, 400MHz, and 800MHz, yielding a total of 3618 sets of subsurface echo data. Each set included echo arrival time, echo amplitude, and echo phase. The propagation velocity in normal media was determined to be 0.12m / ns, and in cavitary media, it was determined to be 0.30m / ns. The direction of signal propagation was vertically downward.

[0031] The subsurface echo data underwent position alignment, time zero-point correction, and frequency band response correction. Before correction, the average position offset at the same detection location across different detection frequency bands was 0.18 m, and the average zero-point offset of the echo time was 0.84 ns. After correction, the average position offset decreased to 0.04 m, and the average zero-point offset of the echo time decreased to 0.09 ns. The corrected subsurface echo data were then arranged according to continuous detection paths, detection locations, and detection frequency bands to generate a multi-frequency subsurface detection image set.

[0032] Continuous reflection ridges were identified in a multi-frequency underground detection image set. The detection location, frequency band, echo arrival time, and echo phase were correlated to form reflection characterization data. Taking the third continuous detection path as an example, 188 of the 201 detection locations had the same depth order, and 186 of the detection locations had the same echo phase change direction. By correlating the spatial location correspondence, depth order correspondence, and echo response correspondence, 117 simultaneously consistent detection locations were identified, forming 351 sets of cross-frequency echo data, which in turn generated cross-frequency homogeneous reflection trajectories.

[0033] Surface fitting was performed on the detection location, detection frequency band, and echo arrival time of 351 sets of transfrequency echo data. The average time difference between the initial propagation time surface and the trajectory time point was 0.47 ns. After six consecutive weighted surface fittings, the maximum time value change between adjacent fitted surfaces gradually decreased from 0.62 ns to 0.03 ns, ultimately generating a propagation time reference surface that reflects the overall changing trend of the subsurface medium interface. Three transfrequency co-source reflection trajectories were formed within the detection area, with average echo arrival times of 18.4 ns, 27.7 ns, and 39.2 ns, corresponding to the three subsurface medium interfaces.

[0034] The actual echo arrival time is compared with the propagation time reference plane to generate propagation time difference sequences corresponding to different detection frequency bands. The 200MHz detection frequency band forms a continuous anomalous section from 4.20m to 5.35m, the 400MHz detection frequency band forms a continuous anomalous section from 4.30m to 5.20m, and the 800MHz detection frequency band forms a continuous anomalous section from 4.35m to 5.15m. There are also a few isolated anomalous positions in the three detection frequency bands, which are removed after continuity processing.

[0035] Within the overlapping intervals of the 200MHz and 400MHz segments, a total of 18 adjacent propagation time advance differences were obtained, of which 17 showed the same direction of change and differed by 0.05m at their turning points. Within the overlapping intervals of the 400MHz and 800MHz segments, a total of 16 adjacent propagation time advance differences were obtained, of which 15 showed the same direction of change and differed by 0.10m at their turning points. Connecting these three segments, which have corresponding lateral positions and consistent propagation time advance trends, sequentially generates a cross-frequency propagation time difference gap chain.

[0036] An anomalous segment was formed in the 200MHz detection band within the range of 7.10m to 7.85m, but no corresponding results were found in the 400MHz and 800MHz detection bands, so it was not connected to the cross-frequency propagation time difference vacancy chain; subsequent pinhole inspection confirmed that the area was a local water-bearing loose area and no underground cavity was formed, indicating that the cross-frequency connection can eliminate anomalous interference in a single detection band.

[0037] Based on the cross-frequency propagation time difference vacancy chain, the lateral extension range from 4.20m to 5.35m was determined. The propagation time advance gradually increased from 0.42ns at the starting position, reaching 1.86ns at the middle position, and then decreased to 0.38ns at the ending position. Combined with the average echo arrival time of 27.7ns and the normal medium propagation velocity of 0.12m / ns, the subsurface medium interface depth of 1.662m was generated, and the vacancy chain spatial characterization data was formed.

[0038] By utilizing the difference between the propagation velocity in normal and void media, the propagation time advance is converted into the equivalent vacancy thickness. A location with a propagation time advance of 1.86 ns corresponds to an equivalent vacancy thickness of 0.186 m, and a location with a propagation time advance of 0.42 ns corresponds to an equivalent vacancy thickness of 0.042 m. Spatial mapping is performed in the opposite direction of the probe signal propagation to generate candidate projections of propagation time difference vacancy corresponding to each continuous probe path and each underground medium interface.

[0039] Six continuous detection paths and three underground medium interfaces form a total of 18 candidate projections for propagation time difference vacancies. After placing the 18 candidate projections in the same underground space, 14 interconnected spatial overlapping parts are formed. By connecting and closing the boundaries of the spatial overlapping parts, a propagation time difference vacancies closed body with a length of 1.08m, a width of 0.46m, and a maximum height of 0.186m is generated.

[0040] The top and bottom boundaries of the propagation time difference vacancy closure were mapped onto six continuous detection paths, resulting in 144 top boundary echo locations and 144 bottom boundary echo locations. By analyzing the echo amplitude and phase, the top and bottom boundaries were corrected. Before correction, the average depths of the top and bottom boundaries were 1.482m and 1.660m, respectively; after correction, they were 1.470m and 1.646m, respectively.

[0041] The final determination of the underground cavity center was at a detection position of 4.78m along the third continuous detection path, with a burial depth of 1.470m and a spatial range of 1.08m in length, 0.46m in width, and 0.18m in height. Small-hole verification confirmed the underground cavity center was located at a detection position of 4.82m, with a top depth of 1.51m, a bottom depth of 1.66m, a length of 1.12m, and a width of 0.44m. Compared with the verification results, the center position difference generated by this invention was 0.04m, the top depth difference was 0.04m, the bottom depth difference was 0.014m, the length difference was 0.04m, and the width difference was 0.02m.

[0042] The above test results show that this invention can eliminate isolated single-frequency anomalies by utilizing multi-frequency echo correspondence and reduce interference caused by underground medium interface undulations by using the propagation time reference surface. Simultaneously, it can convert the propagation time advance into the equivalent thickness of the void and generate the closure range of the underground cavity through the spatial superposition of multiple paths and interfaces, thereby accurately determining the location, depth, and spatial extent of the underground cavity.

[0043] Table 1. Comparison of the overall performance of different underground cavity detection methods

[0044] Based on the false alarm results, the single-frequency manual interpretation method identified 4 suspected anomalies, of which 3 were not confirmed as underground cavities by the pinhole inspection, resulting in a false alarm rate of 75.0%. Although the single-frequency amplitude threshold identification method and the multi-frequency image direct superposition method could eliminate some discrete clutter, their false alarm rates were still 66.7% and 50.0%, respectively. The method of this invention did not retain the false alarm area, mainly because it first arranged the spatial location, depth, and echo response of continuous reflection ridges in different detection frequency bands, and then screened out isolated anomalies that only appeared in a single detection frequency band by cross-frequency propagation time difference vacancy chain.

[0045] In terms of position and depth errors, the method of this invention has a cavity center position error of 0.04m, a top depth error of 0.04m, and a bottom depth error of 0.014m, all of which are superior to the other three methods. This improvement stems from the separation of the propagation time reference surface from the overall undulation of the underground medium interface, which allows the phenomenon of local propagation time advance to be identified separately. At the same time, the top interface response and the bottom interface response further correct the propagation time difference vacancy closure, reducing the deviation caused by relying solely on local echo amplitude to determine the boundary.

[0046] Regarding spatial extent determination, the method of this invention has an average planar dimension error of 0.03m and a boundary integrity rate of 96%, while the direct superposition method using multi-frequency images has errors of 0.06m and 81%, respectively. This invention converts the propagation time advance into the equivalent thickness of the void and places candidate projections of propagation time difference voids corresponding to different underground medium interfaces and adjacent continuous detection paths into the same underground space for superposition and boundary closure, thus forming continuous top, bottom, and lateral boundaries. Therefore, this invention not only reduces false alarms caused by underground pipelines, water-bearing areas, and discrete clutter, but also improves the accuracy of determining the location, depth, and spatial extent of underground cavities.

[0047] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic detection method for underground cavities based on image recognition, characterized in that, Includes the following steps: Multiple continuous detection paths are deployed along the underground area to be detected to determine the detection location. Data is collected using multiple detection frequency bands and corrected to generate a multi-frequency underground detection image group. Identify continuous reflection ridges in multi-frequency underground detection image groups, connect continuous reflection ridges with consistent spatial location, depth arrangement and echo response under different detection frequency bands, and generate cross-frequency homogeneous reflection trajectories and trajectory characterization data; Based on trajectory characterization data, surface fitting is performed on the overall trend of echo arrival time of each cross-frequency co-source reflection trajectory to generate propagation time reference surfaces corresponding to each underground medium interface. The difference between the actual echo arrival time and the propagation time reference plane is calculated to generate a propagation time difference sequence, identify local propagation time advance segments, and determine the propagation time advance amount; Based on cross-frequency homo-source reflection trajectories, local propagation time advance segments with corresponding positions and consistent propagation time advance trends under different detection frequency bands are connected to generate cross-frequency propagation time difference vacancy chains and vacancy chain spatial characterization data. Based on the spatial characterization data of the vacancy chain, the equivalent thickness of the vacancy is generated, and the equivalent thickness of the vacancy is mapped to the underground space. The mapping results of different underground media interfaces and adjacent continuous detection paths are superimposed and the boundaries are closed to generate the propagation time difference vacancy closed body. Based on the echo response correction propagation time difference vacancy closed body corresponding to the top and bottom boundaries, the underground cavity region is generated, and the location, burial depth and spatial range of the underground cavity are determined.

2. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the multi-frequency underground detection image set includes: Multiple continuous detection paths are laid out along the underground area to be detected, and detection positions are determined on each continuous detection path; Multiple detection frequency bands were used to collect underground echo data at each detection location, and the propagation velocity in normal medium, the propagation velocity in cavitary medium, and the propagation direction of the detection signal were obtained. The underground echo data is aligned in position, corrected at zero time, and corrected in frequency band response to generate corrected underground echo data. The corrected underground echo data is then arranged according to the detection location and detection frequency band to generate a multi-frequency underground detection image group.

3. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the cross-frequency homogeneous reflection trajectory and trajectory characterization data includes: By traversing the multi-frequency underground detection image group according to the detection frequency band, and connecting the reflection pixels with continuous depth changes and connected echo phases in adjacent detection locations, a continuous reflection ridge is generated. The continuous reflection ridges are correlated with the detection location, detection frequency band, echo arrival time, and echo phase to generate reflection characterization data. Based on the reflection characterization data, the spatial positions of continuous reflection ridges corresponding to different detection frequency bands are mapped, and spatial position mapping results are generated. Compare the depth arrangement order and echo phase change of the continuous reflection ridges corresponding to different detection frequency bands to generate depth arrangement correspondence results and echo response correspondence results; By connecting continuous reflection ridges with consistent spatial location, depth arrangement, and echo response, cross-frequency homogeneous reflection trajectories are generated, and the corresponding reflection characterization data are collected to generate trajectory characterization data.

4. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the propagation time reference surface includes: The detection location, detection frequency band, and echo arrival time are obtained from the trajectory characterization data, and the trajectory time point set is generated by arranging them according to the detection location and detection frequency band. Using the detection location, detection frequency band, and echo arrival time as the location coordinates, frequency band coordinates, and time coordinates, respectively, an initial surface is fitted to the trajectory time point set to generate an initial propagation time surface; Calculate the surface fitting residual between each trajectory time point and the initial propagation time surface, and compare the fitting residuals of adjacent surfaces along the detection position and detection frequency band to generate position continuity data and frequency band consistency data; By combining location continuity data and frequency band consistency data to determine fitting weights, a weighted surface is fitted to the trajectory time point set, and the surface fitting residuals and fitting weights are updated until the time value changes of two adjacent weighted propagation time surfaces remain stable, thus generating a propagation time trend surface. By associating the propagation time trend surface with the corresponding cross-frequency homologous reflection trajectory, a propagation time reference surface corresponding to the underground medium interface is generated.

5. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The determination of the propagation time advance includes: Based on the detection location and frequency band, the actual echo arrival time is mapped to the propagation time reference plane to determine the corresponding reference echo arrival time, and the difference between the reference echo arrival time and the actual echo arrival time is calculated to generate the propagation time difference value. The propagation time difference values ​​are associated with the detection location and detection frequency band, and arranged along the detection location according to the detection frequency band to generate a propagation time difference sequence; Calculate the time difference change between adjacent detection positions in the propagation time difference sequence, and arrange them along the detection positions to generate a propagation time difference change sequence; By combining the propagation time difference sequence and the propagation time difference change sequence, segments with continuously positive and continuously changing propagation time difference values ​​are identified, isolated positive propagation time difference values ​​appearing at a single detection location are eliminated, the segment start point and segment end point are determined, and local propagation time advance segments are generated. The propagation time difference value corresponding to each detection position in the local propagation time advance segment is determined as the propagation time advance amount, and associated with the corresponding local propagation time advance segment.

6. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the vacancy chain spatial representation data includes: Taking each cross-frequency co-source reflection trajectory as the processing object, local propagation time advance segments are arranged according to the detection frequency band, and the corresponding detection positions and propagation time advance are associated to generate a cross-frequency advance segment set; The local propagation time advance segments corresponding to adjacent detection frequency bands are matched with the same detection position to determine the overlapping range of detection positions between segments and generate horizontal position correspondence data. Arrange the propagation time advance within the overlapping interval of the detection positions according to the detection positions, compare the direction of increase and decrease and the turning point of the propagation time advance in adjacent detection frequency bands, and generate corresponding data on the propagation time advance trend. By combining the data corresponding to the horizontal position and the data corresponding to the propagation time lead trend, the local propagation time lead segments that correspond to the horizontal position and have the same trend of propagation time lead change are connected sequentially to generate the cross-frequency propagation time difference gap chain. Based on the cross-frequency propagation time difference vacancy chain, the lateral extension range and propagation time advance distribution are determined. The depth of the subsurface medium interface is generated by combining the echo arrival time and normal medium propagation velocity in the corresponding cross-frequency co-source reflection trajectory, and then correlated to generate vacancy chain spatial characterization data.

7. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the propagation time difference vacancy closure includes: Arrange the propagation time advance corresponding to each detection position according to the propagation time advance distribution, calculate the time difference per unit propagation distance corresponding to the propagation speed of normal medium and the propagation speed of void medium, and convert the propagation time advance corresponding to each detection position to generate the void equivalent thickness. The reverse mapping starting point corresponding to each detection location is determined based on the depth of the underground medium interface. The equivalent thickness of the vacancy is extended in the opposite direction of the detection signal propagation to determine the reverse mapping endpoint. Within the lateral extension range, the inverse mapping start point and inverse mapping end point corresponding to adjacent detection locations are connected respectively to generate the inverse mapping start boundary and inverse mapping end boundary. The underground space between the two boundaries is then processed to generate a propagation time difference space candidate projection body, which is then associated with the corresponding underground medium interface and continuous detection path. Candidate projections of propagation time difference vacancy corresponding to different underground media interfaces and adjacent continuous detection paths are placed in the same underground space, and overlapping candidate projections of propagation time difference vacancy are connected to generate a candidate projection superposition area. Identify the outer edge of the candidate projection superposition area, generate the outer boundary, connect the continuously distributed outer boundaries between different underground media interfaces and between adjacent continuous detection paths, and close the discontinuous positions to generate a propagation time difference vacancy closure body.

8. The automatic detection method for underground cavities based on image recognition according to claim 1, characterized in that, The generation of the location, depth, and spatial extent of the underground cavity includes: The echo position is determined according to the top and bottom boundaries of the propagation time difference space closed body, the corresponding echo amplitude and echo phase are obtained, and their changes are compared along the detection position to generate the top interface response and bottom interface response. Based on the top and bottom interface responses, the top and bottom interface locations where the echo amplitude changes abruptly and the echo phase changes are identified, respectively. The top and bottom boundaries of the propagation time difference void closure body are corrected by using the top and bottom interface positions respectively, while the lateral boundaries are preserved, to generate underground void regions; Based on the top, bottom, and lateral boundaries of the underground cavity area, the corresponding depth, detection location range, and continuous detection path are determined to generate the location, burial depth, and spatial range of the underground cavity.

9. An automatic underground cavity detection system based on image recognition, comprising the automatic underground cavity detection method based on image recognition as described in any one of claims 1 to 8, characterized in that, include: The image generation module is used to acquire and correct underground echo data and generate multi-frequency underground detection image sets; The trajectory generation module is used to identify and connect continuous reflection ridges corresponding to different detection frequency bands, and generate cross-frequency homogeneous reflection trajectories and trajectory characterization data; The reference surface generation module is used to perform surface fitting on the changing trend of echo arrival time and generate a propagation time reference surface. The advance segment generation module is used to generate local propagation time advance segments and propagation time advance amounts; The vacancy chain generation module is used to connect local propagation time advance segments to generate cross-frequency propagation time difference vacancy chains and vacancy chain spatial representation data; The closed volume generation module is used to generate the equivalent thickness of the void and the candidate projection of the propagation time difference void, and to perform spatial superposition and boundary closure to generate the propagation time difference void closed volume; The cavity detection module is used to correct the top and bottom boundaries of the propagation time difference vacancy closure body, and generate underground cavity areas as well as the location, burial depth and spatial range of underground cavities.

Images