WO2024108971A1

WO2024108971A1 - Agv system for vehicle chassis corrosion evaluation

Info

Publication number: WO2024108971A1
Application number: PCT/CN2023/098465
Authority: WO
Inventors: 王亚飞; 周志松; 邬明宇; 刘旭磊; 李泽星; 张睿韬; 章翼辰
Original assignee: 上海交通大学
Priority date: 2022-11-21
Filing date: 2023-06-06
Publication date: 2024-05-30
Also published as: CN115760058A

Abstract

An AGV system for vehicle chassis corrosion evaluation, wherein particle filtering and image matching are respectively performed by means of a vehicle bottom real-time positioning module according to point cloud information collected by a plane laser scanner and image information collected by a camera, and abnormal value filtering is performed on a positioning result to obtain a vehicle bottom laser and image fusion positioning result; part complexity evaluation, self-adaptive part region segmentation, and corrosion grade evaluation are performed by means of a corrosion evaluation module according to vehicle bottom positioning information and real-time image information collected by the camera so as to obtain a vehicle bottom part corrosion grade evaluation result; and strong edge suppression, local entropy calculation, and neural network determination corrosion treatment are performed by means of a corrosion rechecking module according to part local corrosion image information so as to obtain a part corrosion rechecking result. According to the present invention, detected corrosion regions can be rechecked, false corrosion regions such as red paint-coated parts are screened out, and corrosion grade evaluation accuracy is improved.

Description

AGV system for vehicle chassis corrosion assessment

Technical Field

The present invention relates to a technology in the field of vehicle maintenance, in particular to an AGV system for evaluating vehicle chassis corrosion.

Background technique

Existing chassis corrosion assessment solutions mostly use lifts, ground-fixed detection systems, mobile inspection vehicles, etc. Among them, the mobile inspection vehicle can move freely under the vehicle and collect chassis images through cameras. It has good portability and high detection efficiency. Existing mobile inspection vehicle solutions generally transmit chassis images to maintenance personnel in real time, or splice scattered real-time images into a complete chassis image. It is unable to autonomously detect and evaluate the corrosion area of the collected images, and the degree of automation and intelligence is low.

Summary of the invention

In view of the shortcomings of existing testing devices that are unable to move and have a low degree of automation, the present invention proposes an AGV system for vehicle chassis corrosion assessment, which can autonomously traverse the space under the vehicle, autonomously detect chassis corrosion areas, and autonomously assess chassis corrosion levels. There is no need to prepare a chassis parts data set in advance, and the corrosion level can be accurately assessed based on the corrosion area coverage of the parts. Finally, a chassis corrosion assessment report with reference value is generated, and images of key areas are saved to facilitate maintenance personnel to conduct a comprehensive status assessment of the vehicle chassis. Through a chassis corrosion area review method that integrates a convolutional neural network and an image local entropy value, the detected corrosion areas can be reviewed, pseudo-corrosion areas such as red-painted parts can be screened out, and the accuracy of corrosion level assessment can be improved.

The present invention is achieved through the following technical solutions:

The present invention relates to an AGV system for evaluating the corrosion of a vehicle chassis, comprising: a real-time positioning module for a vehicle bottom, a corrosion evaluation module and a corrosion review module, wherein: the real-time positioning module for the vehicle bottom performs particle filtering and image matching processing respectively according to point cloud information collected by a planar laser scanner and image information collected by a camera, and performs outlier filtering on the positioning result to obtain the vehicle bottom laser and image fusion positioning result; the corrosion evaluation module performs part complexity evaluation, adaptive part area segmentation, and corrosion level evaluation processing according to the vehicle bottom positioning information and the real-time image information collected by the camera to obtain the vehicle bottom part corrosion level evaluation result; the corrosion review module performs strong edge suppression, local entropy calculation and neural network corrosion judgment processing according to the local corrosion image information of the part to obtain the part corrosion review result.

The present invention relates to a vehicle chassis corrosion assessment method based on the above-mentioned AGV system, comprising the following steps:

Step 1: Correct the positioning of the AGV under the vehicle based on the point cloud and image information collected by the AGV;

Step 2: Preprocess the chassis image captured by the AGV-mounted camera;

Step 3: Evaluate the parts complexity of the chassis image;

Step 4: According to the part complexity evaluation result, the chassis image is adaptively segmented into parts regions;

Step 5: Detect the corrosion area of the chassis image based on the color space threshold filtering method and evaluate the corrosion level;

Step 6: Perform strong edge suppression on the eroded area of the chassis image and conduct local entropy corrosion review;

Step 7: Perform neural network corrosion review on the corroded area of the chassis image and fuse the corrosion review results;

Step 8: Add location tags to the corrosion assessment and review results based on the real-time location information of the AGV.

Technical Effects

The present invention uses adaptive part area segmentation technology based on part complexity assessment and local entropy and neural network fusion corrosion review technology based on strong edge suppression. It can adaptively segment parts of different sizes and types based on part complexity assessment, and calculate and evaluate the corrosion levels respectively without preparing part data sets in advance; local entropy and neural network fusion review are performed on the corrosion area to screen out false detection of non-corroded parts that exists in the color space threshold filtering method.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a schematic diagram of the present invention;

FIG2 is a schematic diagram of the appearance corrosion evaluation standard defined in the GMW-15357 document;

FIG3 is a schematic diagram of a simulation scenario;

Figures 4 and 5 are the constructed laser maps, and the red dots are schematic diagrams of real-time point cloud data;

Figures 6 and 7 are schematic diagrams of the corrosion assessment AGV traversing the path;

FIG8 is a schematic diagram of a chassis image after adaptive histogram equalization processing;

FIG9 is a schematic diagram of strong edge detection;

Fig. 10 is a schematic diagram of the local entropy of a part contour;

FIG11 is a schematic diagram of weak edge detection;

FIG12 is a schematic diagram of strong edge suppression;

FIG13 is a schematic diagram of the adaptive Filsenzwab over-segmentation result;

FIG14 is a schematic diagram of the results of merging the adaptive regional adjacency graph;

FIG15 is a schematic diagram of the part region segmentation result after morphological closed region detection;

Figures 16 and 17 are schematic diagrams of HSV color space threshold segmentation of corrosion areas and non-corrosion areas;

FIG18 is a schematic diagram of corrosion assessment results for different parts areas;

FIG19 is a schematic diagram of the segmentation of corrosion sub-regions of a part;

FIG20 is a schematic diagram of local entropy corrosion review of a corrosion sub-region of a part;

FIG21 is a schematic diagram of the neural network corrosion review of the corrosion sub-region of the part;

Figure 22 is a logical diagram of the fusion of local entropy and neural network corrosion review results.

Detailed ways

As shown in Figure 1, this embodiment involves an AGV system for vehicle chassis corrosion assessment, including: a real-time vehicle bottom positioning module, a corrosion assessment module and a corrosion review module, wherein: the real-time vehicle bottom positioning module performs particle filtering and image matching processing according to the point cloud information collected by the planar laser scanner and the image information collected by the camera, and performs outlier filtering on the positioning result to obtain the vehicle bottom laser and image fusion positioning result; the corrosion assessment module performs part complexity assessment, adaptive part area segmentation, and corrosion level assessment processing according to the real-time image information collected by the camera to obtain the vehicle bottom part corrosion level assessment result; the corrosion review module performs strong edge suppression, local entropy calculation and neural network corrosion judgment processing according to the vehicle bottom positioning information and the part local corrosion image information to obtain the part corrosion review result and mark the part corrosion area.

The real-time under-vehicle positioning module includes: a laser map construction unit, a chassis image stitching unit, a map positioning unit and an under-vehicle positioning correction unit, wherein: the laser map construction unit performs particle filtering processing according to the planar laser scanner information to obtain the under-vehicle laser map result; the image map construction unit performs image feature matching and stitching processing according to the image information collected by the camera to obtain the chassis image map result; the map positioning unit performs coordinate system conversion processing according to the relative position information of the vehicle tires in the laser map and the image map to obtain the map conversion matrix result; the under-vehicle positioning correction unit performs outlier filtering processing according to the map conversion matrix and the AGV real-time positioning information to obtain the corrected under-vehicle positioning result.

The corrosion assessment module includes: an image preprocessing unit, a part complexity assessment unit, a part area segmentation unit, a corrosion area segmentation unit and a corrosion level assessment unit, wherein: the image preprocessing unit performs adaptive histogram equalization, downsampling and Gaussian blur processing according to the image information collected by the camera to obtain a preprocessed image result; the part complexity assessment unit performs strong edge detection, local entropy calculation and part complexity calculation processing according to the preprocessed image information to obtain a part complexity result; the part area segmentation unit performs adaptive Filsentzwab segmentation, adaptive area adjacency graph merging and morphological closure detection processing according to the preprocessed image and part complexity information to obtain a part area segmentation result; the corrosion area segmentation unit performs HSV color space filtering processing according to the preprocessed image information to obtain a corrosion area segmentation result; the corrosion level assessment unit performs corrosion coverage calculation and local segmentation processing of the corrosion area according to the part area and corrosion area segmentation information to obtain a corrosion assessment level and a local corrosion image result.

The corrosion review module includes: a local entropy corrosion review unit, a neural network corrosion review unit, and a corrosion area marking unit, wherein: the local entropy corrosion review unit performs strong edge suppression and corrosion entropy calculation processing according to local corrosion image information to obtain a local entropy corrosion review result; the neural network corrosion review unit performs neural network binary classification processing according to local corrosion image information to obtain a neural network corrosion review result; the corrosion area marking unit performs review result fusion and image marking processing according to local entropy and neural network corrosion review information and corrected vehicle bottom positioning information to obtain a marked corrosion area result.

As shown in FIG1 , this embodiment relates to a vehicle chassis corrosion assessment method for the above-mentioned AGV system, comprising the following steps:

Step 1: Correct the positioning of the AGV under the vehicle based on the point cloud and image information collected by the AGV, including:

1.1) Laser map construction: Place the AGV under the vehicle to be detected, and let the AGV randomly explore and construct laser maps under the vehicle. Map, as shown in Figure 3. In this step, since the flat laser scanner of the chassis corrosion assessment AGV is installed at a low height and the distance between vehicle tires is much smaller than the detection range of the laser scanner with a radius of 12 meters, the AGV can scan the tire contour point cloud in real time. By randomly selecting target points in the space under the vehicle, the AGV can autonomously explore the bottom of the vehicle until a complete laser map of the bottom of the vehicle is constructed, as shown in Figures 4-7. In the absence of other obstructions under the vehicle, the laser map of the bottom of the vehicle is expressed as four rectangles representing tires.

In this embodiment, the laser map coordinate system under the vehicle is constructed with the center of the vehicle as the origin, the front of the vehicle as the positive direction of the y-axis, and the rightward direction of the vehicle body as the positive direction of the x-axis. The transposed matrix of the coordinate matrix in the laser coordinate system composed of the four tire centers of the left front tire, the right front tire, the left rear tire, and the right rear tire is

1.2) Image map construction: The AGV simultaneously collects chassis images in real time through a USB camera with a vertical upward viewing angle. Detect SURF feature points in the real-time image, perform FLANN feature matching on the SURF feature points of adjacent frame images, and calculate the homography matrix between adjacent frame images. According to the homography matrix H, adjacent frame images are mapped and transformed and the overlapping parts are spliced, and finally spliced into a complete chassis image map.

In this embodiment, the chassis image map is a pixel coordinate system, the origin is at the upper left corner of the image, the horizontal right direction of the image is the positive direction of the u axis, and the vertical downward direction of the image is the positive direction of the v axis. The transposed matrix of the coordinate matrix in the pixel coordinate system composed of the four tire centers of the left front tire, the right front tire, the left rear tire, and the right rear tire is

1.3) Mapping: Based on the coordinate matrices of the left front tire, right front tire, left rear tire, and right rear tire centers in the laser and pixel coordinate systems, calculate the coordinate conversion matrix between the laser map and the image map Complete map marking.

In this embodiment, for any coordinates under the laser map The coordinates under the image map can be obtained by coordinate conversion Similarly, for any coordinate under the image map The coordinate transformation matrix can be Get its coordinates under the laser map

1.4) Underbody positioning correction: Perform coordinate system conversion and outlier filtering on the image map and laser map positioning results to correct the underbody positioning results.

Since laser map positioning is prone to pose estimation flipping, in this embodiment, the laser map positioning result And the image map positioning result obtained by coordinate system conversion If _xs and _xp are detected to have opposite signs in three consecutive frames, the laser map is flipped along the y-axis. Similarly, if _ys and _yp are detected to have opposite signs in three consecutive frames, the laser map is flipped along the x-axis.

The image map positioning method is prone to mismatching problems. In this embodiment, the image map positioning result And the laser map positioning result obtained by coordinate system conversion If | _up - _us |＞0.1*u _max or | _vp - _vs |＞0.1*v _max , that is, the absolute value of the difference between the positioning results of the image map and the laser map exceeds 10% of the total size of the image map, the positioning result of the image map is considered to be an outlier, and the corrected positioning result is taken as the laser map P _′s .

In this embodiment, for the image map positioning result And the laser map positioning result obtained by coordinate system conversion If | _up - _us |≤0.1* _umax or | _vp - _vs |≤0.1* _vmax , that is, the absolute value of the difference between the positioning results of the image map and the laser map does not exceed 10% of the total size of the image map, then both positioning results are considered normal, and the corrected positioning result is the average of the two.

Step 2: Preprocess the chassis image captured by the AGV-mounted camera, including:

2.1) Adaptive histogram equalization: The original chassis image captured by the camera is subjected to adaptive histogram equalization, as shown in Figure 8. This operation can effectively brighten the insufficiently illuminated area and retain the details of the parts.

In this embodiment, the selected equalization threshold is 0.1.

2.2) Downsampling: Downsample the RGB channels of the chassis image separately. This operation can blur the high-frequency details of the parts and highlight the contour curves of the parts.

In this embodiment, the downsampling ratio selected is

2.3) Gaussian Blur: Perform Gaussian Blur operation on the chassis image. This operation can blur the high-frequency details of the parts and highlight the contour curves of the parts.

In this embodiment, a 5×5 convolution kernel is selected, and the standard deviation σ=0.8.

Step 3: evaluating the parts complexity of the chassis image, specifically comprising:

3.1) Strong edge detection: Perform Canny strong edge detection on the chassis image, as shown in Figure 9. By setting appropriate Gaussian blur standard deviation and link threshold, this operation can detect the strong edges of parts, that is, part contours, and remove those below the link. Weak edge of the threshold.

In this embodiment, the selected Gaussian blur standard deviation σ=3, the minimum link threshold min_val=0.10, and the maximum link threshold max_val=0.20.

3.2) Part complexity evaluation: Based on the image strong edge detection results, the part complexity is evaluated. Generally speaking, the more strong edges there are in the image, that is, the more part contours there are, the higher the part complexity in the image. At the same time, the higher the local entropy value of the image, the more uneven the part distribution is and the higher the part complexity is.

In this embodiment, the image strong edge ratio is calculated based on the number of strong edge pixels _ne and the total number of image pixels _np . The value range is 0≤R _e ≤1.

In this embodiment, in a 5×5 rectangular moving window, according to the local entropy expression The local entropy is calculated for the strong edge detection result, as shown in Figure 10. Since the strong edge detection result is a binary image, the local entropy expression can be simplified to _He = -( _p0 _log2p0 +(1- _p0 ) _log2 (1- _p0 )). By deriving the expression _, we can get When the pixel intensity is 0, the probability p ₀ is When p ₀ is 0 or 1, the local entropy _He _takes the minimum value 0. The local entropy of the image is averaged. The value range is also

In this embodiment, the part complexity evaluation parameter is proposed The value range is 0 ≤ _Ce ≤ 1. The higher the _Ce value, the more complex the parts in the image.

3.3) Adaptive parameter calculation: According to the calculated part complexity evaluation parameter _Ce , the applicable scale parameter scale of the Filsenzwab segmentation algorithm and the applicable merging threshold thresh of the region adjacency graph algorithm are adaptively calculated.

In this embodiment, for images with high part complexity, a low scale parameter is selected. For images with a width of w and a height of h, a scale parameter applicable to the Filsenzwab segmentation algorithm is proposed and calculated. Wherein k is the amplification factor, and 5 is selected in this embodiment.

In this embodiment, for images with high part complexity, a low merging threshold is selected, and a merging threshold suitable for the regional adjacency graph algorithm is proposed and calculated.

Step 4: According to the component complexity evaluation result, the chassis image is adaptively segmented into component regions, specifically comprising:

4.1) Adaptive Filsenzwab segmentation: Based on the adaptively calculated scale parameters The preprocessed chassis image is subjected to adaptive Filsenzwab segmentation to obtain the over-segmentation result of the part area, as shown in Figure 13.

In this embodiment, the minimum segmentation area size min_size=scale is selected.

4.2) Adaptive region adjacency graph merging: Based on the adaptively calculated merging threshold The over-segmentation results of the part region are merged by adaptive region adjacency graph based on color similarity to obtain the preliminary part region segmentation results, as shown in Figure 14.

4.3) Morphological closure detection: Morphological closure detection is performed on part regions that are similar in color but not spatially adjacent, and the part regions are segmented into closed region instances to obtain the final part region segmentation result, as shown in Figure 15.

Step 5: detecting the corrosion area of the chassis image based on the color space threshold filtering method and evaluating the corrosion level, specifically including:

5.1) HSV color space filtering: Convert the image from RGB color space to HSV color space and filter based on a preset threshold.

In this embodiment, for the HSV image, pixels whose H channel values fall within the interval [0, 15] and the interval [170, 180], that is, pixels with red color, are segmented as corrosion area masks, as shown in FIGS. 16 and 17 .

5.2) Evaluation of corrosion level of parts: Perform bitwise AND operation on different parts regions and the corrosion region mask to obtain the number of non-zero pixels e _i , and calculate the overlap rate based on the total number of pixels in the part region n _i As shown in Figure 18.

In this embodiment, the corrosion level of different parts is evaluated according to the corrosion evaluation index based on the corrosion area coverage defined in the GMW-15357 document of General Motors. For example, for a part with a corrosion coverage _Ki = 8%, it is evaluated as level 7 corrosion, that is, slight corrosion.

5.3) Part corrosion sub-region segmentation: The corrosion area whose corrosion coverage exceeds the set threshold is segmented into the minimum rectangular bounding box to facilitate the corrosion review module to review, as shown in Figure 19.

In this embodiment, the corrosion coverage review threshold is set to _Ki ≥ 5%. The corrosion coverage _Ki of the corrosion sub-region of the part shown in FIG19 is 36.85%, which is evaluated as level 6 corrosion, belonging to moderate corrosion.

Step 6: Perform strong edge suppression on the eroded area of the chassis image and conduct local entropy erosion review, specifically including:

6.1) Strong edge suppression: Perform Canny edge detection with different parameters on the corrosion sub-area of the part, and perform subtraction to obtain the corrosion edge map. When the Gaussian blur standard deviation and link threshold are relatively low, Canny edge detection can detect all edges including strong edges of part contours and weak edges of corrosion details, as shown in Figure 11. When the Gaussian blur standard deviation and link threshold are relatively high, Canny edge detection only detects strong edges of part contours, as shown in Figure 9.

In this embodiment, the strong edge of the part contour is expanded with a 3×3 convolution kernel, and the edge detection result is subtracted from the edge detection result obtained under the low link threshold condition to retain the edge map of the part corrosion details, as shown in FIG12 .

In this embodiment, the Gaussian blur standard deviation σ=1, the minimum link threshold min_val=0.05, and the maximum link threshold max_val=0.10 are selected for weak edge detection. The Gaussian blur standard deviation σ=3, the minimum link threshold min_val=0.10, and the maximum link threshold max_val=0.20 are selected for strong edge detection.

6.2) Local entropy corrosion review: The average local entropy is calculated for the corrosion detail edge map of the part, and the corrosion sub-areas with average local entropy greater than the set threshold are reviewed as corrosion occurring, as shown in Figure 20.

In this embodiment, the local entropy expression _He = -( _p0 _log2p0 ₊ (1- _p0 ) _log2 (1- _p0 )), and the corrosion threshold is selected as The average local entropy of the corrosion sub-area of the part shown in Figure 20 is 0.38, and the local entropy corrosion review result is positive, which means that corrosion has occurred.

Step 7: Performing a neural network corrosion review on the corrosion area of the chassis image and fusing the corrosion review results, specifically including:

7.1) Neural network binary classification review: The corrosion/non-corrosion binary classification review of the corrosion sub-area of the part is performed using the pre-trained neural network binary classification model, as shown in FIG21 .

In this embodiment, a VGG16 convolutional neural network is trained according to a metal corrosion data set, a sigmoid function is used as an activation function, and a binary cross entropy is used as a loss function to generate a pre-trained binary classification model.

7.2) Fusion of corrosion review results: Perform a bitwise OR operation on the local entropy and neural network corrosion review results. As long as one of them is judged to be corrosion-existing, the final judgment is that corrosion exists. The specific corrosion judgment logic is shown in Figure 22.

Step 8: Add location tags to the corrosion assessment and review results based on the real-time location information of the AGV, including:

8.1) Corrosion area marking: The corrosion sub-areas of the parts, the corrosion assessment results and the corrosion review results are combined with the corrected vehicle bottom positioning coordinates and marked on the image map as the final corrosion assessment report.

8.2) Save the image of the corroded area: The image of the parts judged to be severely corroded will be saved locally on the AGV for reference by maintenance personnel.

Compared with the existing technology, this method adaptively segments parts of different sizes and types based on part complexity assessment, and calculates and evaluates the corrosion level separately without the need to prepare part data sets in advance; local entropy and neural network fusion review are performed on the corrosion area to screen out the false detection of non-corroded parts that exists in the color space threshold filtering method; based on the joint positioning of laser maps and image maps, it has good versatility for vehicles of different sizes and types and is not restricted by the site.

The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

An AGV system for evaluating vehicle chassis corrosion is characterized in that it includes: a real-time positioning module for the bottom of the vehicle, a corrosion assessment module and a corrosion review module, wherein: the real-time positioning module for the bottom of the vehicle performs particle filtering and image matching processing according to point cloud information collected by a planar laser scanner and image information collected by a camera, and performs outlier filtering on the positioning result to obtain the bottom of the vehicle laser and image fusion positioning result; the corrosion assessment module performs part complexity evaluation, adaptive part area segmentation, and corrosion level evaluation processing according to the real-time image information collected by the camera to obtain the bottom of the vehicle part corrosion level evaluation result; the corrosion review module performs strong edge suppression, local entropy calculation and neural network corrosion judgment processing according to the bottom of the vehicle positioning information and the local corrosion image information of the part to obtain the part corrosion review result and mark the part corrosion area.
According to claim 1, the AGV system for vehicle chassis corrosion assessment is characterized in that the real-time underbody positioning module includes: a laser map construction unit, a chassis image stitching unit, a map positioning unit and an underbody positioning correction unit, wherein: the laser map construction unit performs particle filtering processing according to the planar laser scanner information to obtain the underbody laser map result; the image map construction unit performs image feature matching and stitching processing according to the image information collected by the camera to obtain the chassis image map result; the map positioning unit performs coordinate system conversion processing according to the relative position information of the vehicle tires in the laser map and the image map to obtain the map conversion matrix result; the underbody positioning correction unit performs outlier filtering processing according to the map conversion matrix and the AGV real-time positioning information to obtain the corrected underbody positioning result.
According to claim 1, the AGV system for vehicle chassis corrosion assessment is characterized in that the corrosion assessment module includes: an image preprocessing unit, a part complexity assessment unit, a part area segmentation unit, a corrosion area segmentation unit and a corrosion level assessment unit, wherein: the image preprocessing unit performs adaptive histogram equalization, downsampling, and Gaussian blur processing according to the image information collected by the camera to obtain a preprocessed image result; the part complexity assessment unit performs strong edge detection, local entropy calculation, and part complexity calculation processing according to the preprocessed image information to obtain a part complexity result; the part area segmentation unit performs adaptive Filsentzwab segmentation, adaptive area adjacency graph merging, and morphological closure detection processing according to the preprocessed image and part complexity information to obtain a part area segmentation result; the corrosion area segmentation unit performs HSV color space filtering processing according to the preprocessed image information to obtain a corrosion area segmentation result; the corrosion level assessment unit performs corrosion coverage calculation and local segmentation of the corrosion area according to the part area and corrosion area segmentation information to obtain a corrosion assessment level and a local corrosion image result.
The AGV system for vehicle chassis corrosion assessment according to claim 1 is characterized in that the corrosion review module includes: a local entropy corrosion review unit, a neural network corrosion review unit, and a corrosion area marking unit, wherein: the local entropy corrosion review unit performs strong edge suppression and corrosion entropy calculation processing according to local corrosion image information to obtain local Entropy corrosion review results; the neural network corrosion review unit performs neural network binary classification processing based on the local corrosion image information to obtain the neural network corrosion review results; the corrosion area marking unit performs review result fusion and image marking processing based on the local entropy and neural network corrosion review information and the corrected vehicle bottom positioning information to obtain the marked corrosion area results.
A method for evaluating vehicle chassis corrosion according to any one of claims 1 to 4, characterized in that it comprises the following steps:

Step 1: Correct the positioning of the AGV under the vehicle based on the point cloud and image information collected by the AGV;

Step 2: Preprocess the chassis image captured by the AGV-mounted camera;

Step 3: Evaluate the parts complexity of the chassis image;

Step 4: According to the part complexity evaluation result, the chassis image is adaptively segmented into parts regions;

Step 5: Detect the corrosion area of the chassis image based on the color space threshold filtering method and evaluate the corrosion level;

Step 6: Perform strong edge suppression on the eroded area of the chassis image and conduct local entropy corrosion review;

Step 7: Perform neural network corrosion review on the corroded area of the chassis image and fuse the corrosion review results;

Step 8: Add location tags to the corrosion assessment and review results based on the real-time location information of the AGV.
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 1 comprises:

1.1) Laser map construction: The AGV is placed under the vehicle to be inspected, and the AGV randomly explores and constructs a laser map under the vehicle. The AGV scans the tire contour point cloud in real time and obtains a laser map of the vehicle bottom containing four rectangles representing the tires.

1.2) Image map construction: The AGV simultaneously collects chassis images in real time through a USB camera with a vertical upward viewing angle; detects SURF feature points in the real-time image, performs FLANN feature matching on the SURF feature points of adjacent frame images, and calculates the homography matrix between adjacent frame images. According to the homography matrix H, the adjacent frame images are mapped and transformed and the overlapping parts are spliced, and finally spliced into a complete chassis image map;

1.3) Mapping: Based on the coordinate matrices of the left front tire, right front tire, left rear tire, and right rear tire centers in the laser and pixel coordinate systems, calculate the coordinate conversion matrix between the laser map and the image map Complete map marking;

1.4) Underbody positioning correction: Perform coordinate system conversion and outlier filtering on the image map and laser map positioning results to correct the underbody positioning results.
The vehicle chassis corrosion assessment method according to claim 5, characterized in that the step 2 comprises:

2.1) Adaptive histogram equalization: Adaptive histogram equalization is performed on the original chassis image captured by the camera;

2.2) Downsampling: Downsample the RGB channels of the chassis image separately; this operation can blur the high-frequency details of the parts and highlight the contour curves of the parts;

2.3) Gaussian blur: Perform Gaussian blur operation on the chassis image; this operation can blur the high-frequency details of the parts and highlight the contour curves of the parts.
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 3 comprises:

3.1) Strong edge detection: Perform Canny strong edge detection on the chassis image;

3.2) Part complexity evaluation: Based on the image strong edge detection results, the part complexity is evaluated, that is, the image strong edge ratio is calculated based on the number of strong edge pixels ne and the total number of image pixels np . The value range is 0≤R e ≤1;

3.3) Adaptive parameter calculation: According to the calculated part complexity evaluation parameter Ce , adaptively calculate the scale parameter scale applicable to the Filsenzwab segmentation algorithm and the merging threshold thresh applicable to the region adjacency graph algorithm;

The Filsenzwab segmentation algorithm is applicable to the scale parameter Where k is the magnification factor,

The merge threshold
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 4 comprises:

4.1) Adaptive Filsenzwab segmentation: Based on the adaptively calculated scale parameters Perform adaptive Filsenzwab segmentation on the preprocessed chassis image to obtain the over-segmentation result of the part area;

4.2) Adaptive region adjacency graph merging: Based on the adaptively calculated merging threshold The over-segmentation results of the part region are merged by adaptive region adjacency graph based on color similarity to obtain the preliminary part region segmentation results;

4.3) Morphological closure detection: Perform morphological closure detection on part regions that are similar in color but not spatially adjacent, segment the part regions into closed region instances, and obtain the final part region segmentation result.
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 5 comprises:

5.1) HSV color space filtering: convert the image from RGB color space to HSV color space and filter based on a preset threshold;

5.2) Evaluation of corrosion level of parts: Perform bitwise AND operation on different parts regions and the corrosion region mask to obtain the number of non-zero pixels e i , and calculate the overlap rate based on the total number of pixels in the part region n i

5.3) Part corrosion sub-region segmentation: The corrosion area whose corrosion coverage exceeds the set threshold is segmented into the minimum rectangular bounding box to facilitate the corrosion review module to review.
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 6 comprises:

6.1) Strong edge suppression: Perform Canny edge detection with different parameters on the corrosion sub-area of the part, and perform subtraction to obtain the corrosion edge map;

6.2) Local entropy corrosion review: The average local entropy is calculated for the edge map of the corrosion details of the parts, and the corrosion sub-areas with average local entropy greater than the set threshold are reviewed as corrosion;

The local entropy expression He = -(p 0 log 2 p 0 +(1-p 0 )log 2 (1-p 0 )), the corrosion threshold is selected as
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 7 comprises:

7.1) Neural network binary classification review: The corrosion/non-corrosion binary classification review of the corrosion sub-area of the parts is performed through the pre-trained neural network binary classification model;

The neural network binary classification model is a VGG16 convolutional neural network, the sigmoid function is an activation function, and the binary cross entropy is a loss function;

7.2) Fusion of corrosion review results: Perform a bitwise OR operation on the local entropy and neural network corrosion review results. As long as one of them is judged to be corroded, the final judgment is that corrosion exists.
The vehicle chassis corrosion assessment method according to claim 5 is characterized in that the step 8 comprises:

8.1) Corrosion area marking: The corrosion sub-areas of the parts, the corrosion assessment results and the corrosion review results are combined with the corrected vehicle bottom positioning coordinates and marked on the image map as the final corrosion assessment report;

8.2) Save the image of the corroded area: The image of the parts judged to be severely corroded will be saved locally on the AGV for reference by maintenance personnel.