Method for finding and identifying path diseases by using persistent coherence

Abstract

The invention discloses a method for searching and identifying road diseases by using persistent coherence, and relates to the technical field of road defect detection. According to a topological data analysis method based on persistent coherence, path finding is carried out by using persistent coherence. Comprising the following steps: step 1, zooming an image again; 2, finding out a topological structure by using persistent coherence; step 3, bonding the area near the position; and 4, analyzing the shapes and sizes of different segmented regions to detect and identify the road disease. According to the topological fingerprint generation method based on continuous coherence, a novel and powerful technical path is provided for solving the essential feature extraction problem of a two-dimensional object, and path diseases can be found and recognized more accurately.

Description

Technical Field

[0001] This invention relates to the field of road defect detection technology. Background Technology

[0002] Roads develop various types of damage over time due to use and natural factors. These defects not only affect driving comfort but also threaten traffic safety. Below are some common road defects and their descriptions: I. Cracks

[0003] Cracks are one of the most common and earliest defects in asphalt pavements. Based on their morphology and causes, they are mainly classified into longitudinal cracks, transverse cracks, and network cracks (also known as tortoise shell cracks or block cracks). Longitudinal cracks run roughly parallel to the road centerline and are elongated. They are primarily caused by uneven settlement of the subgrade or base course, improper construction joint treatment, or differences in compaction at the junction of the old and new subgrades during road widening. Additionally, material shrinkage due to temperature changes can also cause longitudinal cracks. Transverse cracks run roughly perpendicular to the road centerline and typically penetrate the entire pavement or a portion thereof. They are mainly temperature shrinkage cracks. As temperatures drop, the asphalt surface layer and base course shrink, generating significant tensile stress. When this stress exceeds the material's tensile strength, the pavement cracks. These cracks are usually regular and evenly spaced. Network cracks are a series of intersecting cracks that divide the pavement into polygonal, tortoise shell-like or net-like fragments. Based on fragment size, they can be classified as fine network cracks or wide network cracks. Network cracks are a typical manifestation of fatigue damage to pavement structures. They are usually caused by insufficient strength of pavement structural layers (such as base courses). Under repeated traffic loads, fatigue failure occurs, and network cracks appear in local areas, which means that the structural bearing capacity of the area has been severely reduced.

[0004] II. Deformation type.

[0005] These types of defects mainly manifest as changes in road surface shape, affecting driving stability, including excessively deep ruts, excessively high bumps, and excessively deep subsidence. Ruts are longitudinal, band-shaped grooves appearing on the wheel tracks of the roadway; in cross-section, the road surface appears as if two trenches have sunk. They can be classified into compacted ruts, fluid ruts, and structural ruts. Bumps are bulges formed by localized uplifts in the road surface, usually occurring alongside ruts. They are caused by the asphalt surface layer undergoing shear flow and accumulation under the combined action of horizontal forces from wheels (such as braking and starting) and high temperatures. Subsidence is a depression formed after significant subsidence in a localized area of ​​the road surface. The main causes are uneven soil composition in the subgrade, insufficient compaction, or the hollowing out of the subgrade due to underground pipe leaks or excavation, leading to uneven settlement.

[0006] III. Loose and pitted types.

[0007] These types of defects manifest directly as the loss of pavement materials, forming cavities. Potholes are various shapes of depressions formed after localized loss of pavement materials, typically deeper than 2 cm. They are usually the result of further deterioration of other defects (such as cracks and loosening), and potholes pose an immediate threat to driving safety. Loosening refers to the shedding of surface aggregate particles, with both coarse and fine aggregates being lost. The main causes are poor adhesion between asphalt and aggregates (stripping), insufficient asphalt content in the asphalt mixture, or asphalt aging.

[0008] IV. Surface damage category.

[0009] This type of damage can be divided into bleeding and polishing. Bleeding occurs when a layer of free asphalt appears on the road surface during hot seasons, turning the road surface into a smooth, reflective black mirror. Polishing occurs when the sharp edges of the surface aggregate are worn down by vehicle wheels, resulting in a smooth road surface and a significant decrease in skid resistance.

[0010] The above-mentioned common road defects are often interconnected. For example, cracks, if not treated promptly, can develop into potholes; ruts easily accumulate water, accelerating road surface damage. Therefore, timely detection and implementation of correct maintenance measures are crucial. Traditional methods for detecting road defects rely on visual inspection; however, this method has several drawbacks: First, the process is highly subjective, making it difficult to standardize criteria. Different inspectors may have different standards for defining, judging the severity and type of defects. For example, one person might classify the same crack as "moderate," while another might classify it as "severe." The accuracy of the inspection results highly depends on the inspector's personal experience and professional level. Novices are prone to overlooking early, subtle defects or misjudging complex defects. Visual inspection typically uses qualitative descriptions such as "light," "moderate," and "severe," lacking precise quantitative data (such as the exact width of cracks, the depth of ruts in millimeters, and the specific volume of potholes). This makes it difficult to assess the severity of defects, prepare maintenance budgets, and select the best treatment plan.

[0011] Secondly, it is inefficient and costly. To ensure no areas are missed, manual inspections typically need to be conducted at low speeds or with some lanes closed, which severely impacts inspection efficiency and makes it difficult to quickly complete large-scale road network surveys. A large number of skilled technicians are required for fieldwork, and the accumulated costs of their wages, training, transportation, and equipment are considerable. Low-speed inspections also occupy road resources, potentially causing traffic congestion and posing safety risks to inspection personnel and other road users.

[0012] Third, the safety is poor and the risks are high; testing personnel need to work on the roadway or shoulder, and even with safety measures in place, they still face a high risk of being hit by vehicles. They also need to be exposed to the outdoors for extended periods, facing adverse environmental factors such as weather, noise, and air pollution, which poses a challenge to their health.

[0013] Fourth, data management is rudimentary and lacks historical traceability. Traditional inspections rely heavily on paper-and-pen records, photographs, and manual location markings, which are prone to errors and omissions. The non-digital recording format makes it difficult to input data into a database, hindering efficient querying, statistical analysis, and in-depth analysis. Furthermore, it makes it difficult to accurately compare data with historical data to track disease development trends. Manually recorded location information (such as "100 meters east of XX intersection") is inaccurate, creating difficulties for subsequent precise maintenance work.

[0014] Fifth, the coverage is low, and it is easy to miss some defects. The human eye is prone to fatigue, and in long-term, repetitive work, it is inevitable that some minor or hidden defects (such as early-stage micro-cracks or slight loosening) will be missed. For dangerous or hard-to-reach areas such as the median strip of highways, the sides of viaducts, and the top of tunnels, manual inspection often cannot effectively cover them, forming blind spots.

[0015] Sixth, there is a lack of decision support for preventative maintenance; the naked eye can usually only detect damage when it has developed to a visible extent, at which point the best opportunity for preventative maintenance has often been missed. The naked eye can only observe surface phenomena and is powerless to detect damage inside the pavement structure (such as loose base layer and hidden voids), which are often the root cause of surface damage.

[0016] With the increasing demands for road maintenance and management, traditional manual inspection methods are no longer sufficient to meet the requirements of efficient, accurate, and large-scale detection. In recent years, deep learning-based target detection technologies, especially the YOLO (YouOnlyLookOnce) series of algorithms, have provided powerful technical means to solve this problem and are leading road defect detection into a new stage of intelligence. However, despite YOLO's obvious advantages, its application still faces some challenges, such as: the need for a large amount of high-quality labeled data; the need to continuously improve robustness to occlusion, changes in lighting, and complex backgrounds; and how to more deeply integrate the detection results with the maintenance decision-making system. Summary of the Invention

[0017] To address the above problems, this invention proposes a method for finding and identifying path defects using persistent homology, based on a topology data analysis method using persistent homology to find path defects.

[0018] The technical solution of the present invention includes the following steps: Step 1: Rescale the image; A three-dimensional image of the road surface is obtained using laser scanning (each point in the two-dimensional image has a road surface height value), and the image is first reduced by convolution.

[0019] Step 2: Use persistent homology to find the topology; The standard persistent cohomology method is used to find the topological structure of the three-dimensional image, that is, which points on the road are selected, with high values ​​of 1 and low values ​​of 0. This characterization records the topological structure of the road. The calculation is performed on 3D image slices, considering only H0 (0-dimensional homology) and H1 (1-dimensional homology).

[0020] H0 represents the continuous cohomology of connected components; The longer the "lifetime" (the threshold interval from birth to death) of an H0 component, the more significant and stable the connected region it represents.

[0021] H1 represents the 1D topological feature of the image, also known as a "hole" or "loop". In a 3D image, a hole is a dark area (such as black) that is completely surrounded by a bright area (such as white).

[0022] The longer the "lifetime" of an H1 component, the more robust and significant the hole is, requiring a high threshold (a very bright pixel) to fill it.

[0023] Step 3: Adhere the areas near the location; Since a path defect may consist of multiple topologies, i.e., covered by multiple characterization parts in step 2, a depth threshold is set for locations with a large number of persistent cohomologies that are close to each other. Only regions with persistence values ​​greater than the depth threshold are considered, and then the regions at these locations are glued together to characterize the path defect. Step 4: Analyze the shape and size of different segmented regions to detect and identify road defects; Different road defects have different topological properties, while the persistence graphs of the same defect are similar to each other. The persistence coherence of different types of defects is analyzed and used as a fingerprint of the road. Step 5: After obtaining fingerprints, different categories can be identified, and then the corresponding pathological defects for each group can be determined. Considering the characterization of all regions near the adhesive location in step 3 and their continuous coherence information, these associations are identified using an unsupervised artificial intelligence model. Then, in a new image, the associations learned by this unsupervised artificial intelligence are used to determine which type of pathology it is.

[0024] The topological data analysis method based on persistent homology of this invention can capture the invariant topological features of two-dimensional objects at multiple scales from a topological perspective and generate a standardized "fingerprint" called a "persistent graph" or "barcode graph." This fingerprint quantifies the generation and disappearance processes of topological structures such as connected components and holes of the object, and has high robustness and distinguishability. Using this persistent graph as the "fingerprint" of the object has the following significant advantages, making it extremely valuable in the application scenarios targeted by this invention.

[0025] 1. Deformation robustness: It is not sensitive to translation, rotation, scaling and continuous deformation of the object because these operations do not change its topology.

[0026] 2. Noise stability: Transient, noise-induced topological features are concentrated near the diagonal due to their short persistence, making them easy to identify and filter.

[0027] III. Information richness and computability: This fingerprint transforms complex shape information into a structured set of points (persistent graph), which can be directly used for vectorization processing and as input to machine learning models to achieve efficient object classification, retrieval and comparison.

[0028] In summary, the topological fingerprint generation method based on continuous homology provides a novel and powerful technical approach to solving the problem of essential feature extraction of two-dimensional objects, and can more accurately find and identify path defects. Detailed Implementation

[0029] Figure 1 This is an example of the 3D road surface image obtained in step 1; Figure 2 This is a diagram illustrating the calculation process in step 2. Detailed Implementation

[0030] To clearly illustrate the technical features of the present invention, the present invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings.

[0031] This case presents an alternative to YOLO, employing the technique of continuous coherence. The core idea of ​​continuous coherence is that it doesn't treat an object as a static geometric whole, but rather observes the evolution of its topology through a progressively increasing scale parameter (such as a distance threshold or density radius). For a two-dimensional object (which can be viewed as a set of two-dimensional points or a binary image), this process is typically achieved by constructing a filter, specifically: Data construction: Representing a two-dimensional object as a set of data points. These points can be edge pixels of an image, contour points, or a set of points obtained through specific sampling rules.

[0032] Filtering construction: Select a scale parameter ε and draw a disk with a radius of ε around each data point. As ε gradually increases from 0, these disks will begin to overlap and merge, forming a complex connected structure. This series of geometric complexes that evolve with scale constitutes a filtering process.

[0033] During this filtering process, the "lifecycle" of topological features is continuously tracked in a coherent manner. For two-dimensional objects, the focus is mainly on 0-dimensional and 1-dimensional features.

[0034] From the perspective of 0-dimensional features (connected components): Birth: When the scale ε reaches a certain value, an isolated point or a group of points forms an independent connected component because its disk is connected for the first time. This feature is called "birth".

[0035] Death: When the disk of a connected component is merged with the disk of an earlier "born" connected component, it "dies"—because it is merged into a more "old" component.

[0036] Durability: The durability of a feature is the difference between its "death scale" and its "birth scale". The longer the durability of a feature, the more it is considered to represent the essential structure of the object.

[0037] From the perspective of 1-dimensional features (holes): Birth: A one-dimensional void (i.e., a ring) is "born" when a set of interconnected disks forms a closed ring, and the inner region of the ring is not covered by any disk.

[0038] Death: The cavity "dies" when the scale ε increases to a point where a disk covers and "fills" the center of the cavity.

[0039] Durability: Also calculated from the difference between the "death scale" and the "birth scale". Large, significant voids survive longer than small, noise-generated voids.

[0040] Ultimately, the (birth, death) coordinate pairs of all topological features constitute the topological fingerprint of the object, which is usually visualized in the form of a persistent graph. In this graph, each point represents a topological feature, with its x-axis representing the birth scale and its y-axis representing the death scale. The farther a point is from the diagonal, the longer its persistence and the more representative it is.

[0041] The method for finding and identifying path defects using persistent homology in this invention includes the following steps: Step 1: Rescale the image; Using laser scanning to obtain a 3D image of the road surface (each point in the 2D image has a road surface height value) (e.g.) Figure 1 The original image resolution is about 1 millimeter, which is too high for detecting road defects. Therefore, the image needs to be reduced in size by convolution.

[0042] Step 2: Use persistent homology to find the topology; In this case, using a public program (such as crisper) (reference: https: / / pypi.org / project / cripser / ), standard persistent cohomology methods can be used to find the "characterized part" of the topology of the 3D image (i.e., which points on the path are selected, such as...). Figure 2The convex surface is obtained using a standard persistent cohomology procedure, with a value of 1 for high values ​​and 0 for low values. This "characterized part" records the topological structure of this pathology. The calculation is performed on 3D image slices, considering only H0 (0-dimensional homology) and H1 (1-dimensional homology).

[0043] H0 represents the continuous cohomology of connected components; The longer the "lifetime" (the threshold interval from birth to death) of an H0 component, the more significant and stable the connected region it represents.

[0044] H1 represents the 1D topological feature of the image, also known as a "hole" or "loop". In a 3D image, a hole is a dark area (such as black) that is completely surrounded by a bright area (such as white).

[0045] The longer the "lifetime" of an H1 component, the more robust and significant the hole is, requiring a high threshold (a very bright pixel) to fill it.

[0046] Step 3: Adhere the areas near the location; Since a path defect may consist of multiple topological structures (i.e., the path defect is covered by multiple "characterization parts" in step 2, considering only one "characterization part" would only yield partial information), a depth threshold (i.e., the lower limit of the depth difference between the disappearance and appearance of homology) is set for locations with a large number of persistent homology points that are close together. Only regions with persistence values ​​greater than the depth threshold are considered, because the lifetime of points caused by noise is relatively short, and these are likely to be the locations where path defects occur. Then, the nodes are joined (i.e., in step 2, for example, if all different characterization parts of the path defect are 0, we combine all the "characterization parts" covering the path defect that are 0, i.e., all points in the two-dimensional space where any "characterization part" is 0). This region is the characterization of the path defect.

[0047] Step 4: Analyze the shape and size of different segmented regions to detect and identify road defects; Different road defects are expected to have different topological properties, while the persistence maps of the same defect are similar to each other. Therefore, this patent aims to analyze the persistence coherence of different types of defects and use it as a "fingerprint" of the road.

[0048] Step 5: After obtaining the "fingerprints", different categories can be distinguished. Then, the corresponding pathological disease for each group is determined (if the method is effective, only a few samples should be needed for classification).

[0049] The specific method is as follows: Consider the "characterized portion" of all regions near the adhesion location in step 3 (this is used to distinguish the location of the path defect) and its continuous coherence information (such as birth location, death location, and "lifespan"). This information is specifically related to the path defect, but it is difficult to see with the naked eye. These relationships can be identified using an unsupervised artificial intelligence model, and then, in a new image, the type of path defect can be identified using the relationships learned by this unsupervised artificial intelligence (see...). Figure 2 ).

[0050] There are many specific ways to implement this invention. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A method for finding and identifying road defects using persistent homology, characterized in that, Includes the following steps: Step 1: Rescale the image; Laser scanning is used to obtain 3D images of the road surface, and the images are first reduced in size by convolution. Step 2: Use persistent homology to find the topology; The calculations for 3D image slices only consider 0-dimensional homology and 1-dimensional homology. The standard persistent cohomology method is used to find the topological structure of the three-dimensional image, that is, which points on the road are selected, with high values ​​of 1 and low values ​​of 0. This characterization records the topological structure of the road. Step 3: Adhere the areas near the location; Since a path defect may consist of multiple topologies, i.e., covered by multiple characterization parts in step 2, a depth threshold is set for locations with a large number of persistent cohomologies that are close to each other. Only regions with persistence values ​​greater than the depth threshold are considered, and then the regions at these locations are glued together to characterize the path defect. Step 4: Analyze the shape and size of different segmented regions to detect and identify road defects; Different road defects have different topological properties, while the persistence graphs of the same defect are similar to each other. The persistence coherence of different types of defects is analyzed and used as a fingerprint of the road. Step 5: After obtaining fingerprints, different categories can be identified, and then the corresponding pathological defects for each group can be determined. Considering the characterization of all regions near the adhesive location in step 3 and their continuous coherence information, these associations are identified using an unsupervised artificial intelligence model. Then, in a new image, the associations learned by this unsupervised artificial intelligence are used to determine which type of pathology it is.

Images