Wheeled robot trafficability determination algorithm under road surface defect condition

Abstract

The invention relates to the technical field of computer vision and robot environment perception, in particular to a wheeled robot trafficability judgment algorithm under a pavement defect condition, which comprises the following steps: S1, designing a wheeled robot hardware system which comprises a carried camera for collecting pavement images in real time; s2, constructing an image data set containing a plurality of pavement defects, and marking the image data set; s3, based on a YOLO11n-seg model, developing a YOLO11n-RVD target detection segmentation model, and carrying out pavement defect identification and segmentation on the input image; s4, tracking the identified defects by using a multi-target tracking algorithm, and outputting a trafficability risk level in combination with a risk judgment algorithm; the problems of light weight of a target detection segmentation model and small target detection and positioning in the prior art are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and robot environment perception, and particularly relates to a wheel robot passability determination algorithm under road defect conditions. BACKGROUND

[0002] In the process of outdoor autonomous navigation and inspection of a wheel robot, its driving safety depends not only on successful obstacle avoidance of convex obstacles, but also on effective perception and passability determination of road defects. However, current visual navigation algorithms of robots mainly focus on the identification and detection of convex obstacles, and there is a significant blind area in the perception of road defects. This blind area makes the robot easily trapped, which leads to task interruption, mechanism damage and even the risk of overturning. Therefore, developing a detection algorithm that can accurately identify, segment and quantitatively evaluate road defects is crucial to ensure the autonomy and reliability of the wheel robot.

[0003] At present, the detection of road defects in complex road environments still faces severe challenges. First, the variability of complex road environments and road defects themselves constitutes the main technical bottleneck. Road defects are mostly in the form of multi-scale cracks and potholes, and their visibility is affected by various factors: common water reflection, dynamic shadow, oil pollution and uneven wear, etc. Environmental interference leads to poor robustness and high miss rate of traditional detection algorithms; at the same time, defects differ significantly in scale and shape, from fine defects to wide net-like defects, with blurred edges and irregular shapes, and are easily obscured by fallen leaves and silt, making it difficult for existing models to meet the requirements of actual application in terms of recognition accuracy and positioning accuracy of small target defects.

[0004] Existing research attempts to improve deep learning models to address these challenges, but still has limitations. For example: Mingming Zhang et al. proposed YOLO11-Pear in (Zhang M, Ye S, Zhao S, et al. Pear Object Detection in Complex Orchard Environment Based on Improved YOLO11[J]. Symmetry, 2025, 17(2): 255-255.) which introduced C2PSS to replace C2PSA module to improve the detection ability of the model for obscured pears, added a small target detection layer to improve the detection ability of the model for small pears, and replaced the up-sampling process with DySample to improve the processing speed and expand the application range of the model. However, the calculation amount is too large after adding the small target detection layer, and the post-processing is more time-consuming, which does not meet the lightweight factor.

[0005] Zhao Xiaozhuo et al. (Zhao Xiaozhuo, Zhang Chengtao, Xu Jihuo, et al. Obstacle detection research based on OSTD-YOLO complex road conditions. Journal of Sensor Technology and Application. 2025 Apr 30; 13: 229.) proposed an improved OSTD-YOLO model, constructed a feature pyramid module EMTFP to enhance small targets, and introduced a DyHead attention mechanism detection head to realize adaptive feature enhancement in space, scale and task dimension. However, the model has a large number of parameters and a large model size, which is not suitable for use on mobile devices.

[0006] Patent CN202411850101.8 (application date December 16, 2024, publication date January 14, 2025) discloses a coarse-grained soil particle tracking and crushing matching method based on computer vision. The method adds a CBAM attention mechanism to the backbone part of the YOLO11 model and modifies the CIoU to Focal-EIoU to improve the recognition accuracy of the model. BoT-SORT, PIV algorithm, SIFT algorithm, and FLANN are used for tracking, feature point extraction, and matching to obtain soil particle crushing matching results. However, the introduction of the CBAM module and various algorithms increases the computational complexity of the model, which slows down the inference speed and affects real-time performance.

[0007] Patent CN202510117873.9 (application date January 24, 2025, publication date March 4, 2025) discloses a load tap changer fault diagnosis method based on lightweight YOLO11. StarNet is used to replace the YOLO model backbone network, and star operation is used to enhance feature extraction capability while maintaining network compactness. An efficient joint convolution detection head EJCD is designed to realize model lightweight and multi-scale target detection through grouped normalization and convolution parameter sharing. However, the inference speed is accelerated due to the reduction of model complexity, and the detection effect of smaller objects in the image is not well improved.

[0008] In summary, the existing road pavement defect detection and recognition algorithms have the following limitations:

[0009] 1. In complex environmental interference such as water reflection, dynamic shadow, and oil pollution, the algorithm has poor robustness, poor detection integrity for different scale defects, high miss detection rate, and limited geometric measurement accuracy.

[0010] 2. The complex model structure introduced to improve the sensitivity of multi-scale defects significantly increases the computational cost and parameter quantity, making it difficult to achieve efficient deployment and real-time inference in resource-constrained mobile robot embedded systems.

[0011] Therefore, the present application provides a wheeled robot passability determination algorithm under road defect conditions to solve the above problems. SUMMARY

[0012] In order to solve the problems of lightweight and small target detection positioning of the target detection segmentation model in the prior art, the present application provides a wheeled robot passability determination algorithm under road defect conditions.

[0013] The present application is implemented by adopting the following technical solutions:

[0014] A wheeled robot passability determination algorithm under road defect conditions comprises the following steps:

[0015] S1: Design a wheeled robot hardware system, including a camera for real-time collection of road images;

[0016] S2: Construct an image dataset containing various road defects and label them;

[0017] S3: Based on the YOLO11n-seg model, develop a YOLO11n-RVD target detection segmentation model to identify and segment road defects for input images; wherein the YOLO11n-RVD target detection segmentation model introduces a starnet structure and a C2DIC module in the backbone network, introduces an EP small target detection layer and a Star Blocks in the neck network, and introduces an SGCLDH lightweight detection head in the head network;

[0018] S4: Use a multi-target tracking algorithm to track the identified defects, and output the passability risk level in combination with a risk determination algorithm.

[0019] Further, in step S3, the backbone network adopts a starnet_s050 structure and replaces the original C2PSA module with a C2DIC module; the C2DIC module adopts a two-stage spatial attention TSSA mechanism, first calculates local attention in a local window to capture fine-grained details, then groups the feature maps through a grouping strategy and calculates global attention within the group to model long-distance dependencies, and finally adds the outputs of local attention and global attention to realize feature complementation; wherein the grouping strategy groups the channel dimensions of keys K and values V, reduces the spatial complexity of attention calculation from O(H²W²) to O(HW(M + G²)), where H and W are the height and width of the feature map, M is the local window size, and G is the number of groups.

[0020] Further, in step S3, the EP small target detection layer is improved based on the PAFPN structure, and the P2 feature layer is fused with the P3 feature layer after SPDConv slicing operation to enhance the small target feature reservation and extraction capability. The slicing operation is realized by the following formula:

[0021] ;

[0022] ;

[0023] In the formula, is the input feature map from the P2 feature layer, is the input feature map from the P3 feature layer, represents a feature splicing operation, is the output feature map of the enhanced P3 feature layer after fusion.

[0024] Further, the EP small target detection layer comprises an SEPOKM structure, which integrates the OKM branch and the Initial branch, and integrates the frequency guide modulation module and the multi-scale convolution mechanism to improve the detection accuracy of multi-scale defects.

[0025] Further, in step S3, the SGCLDH lightweight detection head adopts an SGConv module, which generates equivalent convolution kernels by fusing the weights and biases of multiple branch convolutions to realize efficient feature enhancement. The process is defined by the following formula:

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, , are the weight matrix and bias term of the center difference convolution, horizontal difference convolution, vertical difference convolution, angle difference convolution and standard convolution, respectively, is the input feature map, is the convolution operation, is the fused weight matrix, is the fused bias term, is the output feature map after convolution calculation.

[0030] Further, the output of the SGConv module is processed by group normalization and SiLU activation function in turn:

[0031] ;

[0032] ;

[0033] In the formula, GroupNorm 16 represents normalizing the channel by 16 groups.

[0034] Further, in step S4, the BoT-SORT algorithm is used to track the identified defects, and the position, ID and motion trajectory information of the defects are obtained.

[0035] Further, in step S4, the risk determination algorithm sets three risk threshold values by calculating the ratio L of the width of the widest part of the defect to the static tire ground contact width , i.e. the road defect-tire width ratio.

[0036] Further, in step S2, the image data set contains no less than 5000 road defect images, which are divided into training set, validation set and test set in the ratio of 7:1:2.

[0037] Further, the passability determination algorithm of the wheeled robot further comprises a graphical interface developed based on PyQt5, which is used to display the defect detection result, tracking trajectory, risk level and robot motion control state in real time, supports user interaction and data recording.

[0038] The present application provides a passability determination algorithm of a wheeled robot under road defect conditions, which has the following advantages compared with the prior art:

[0039] Firstly, the YOLO11n-RVD target detection and segmentation model can accurately identify road defects and reduce the size of model parameters, and speed up the inference speed.

[0040] Secondly, the algorithm integrates a multi-scale defect detection network, which effectively improves the recognition accuracy and accuracy.

[0041] Thirdly, according to the defect positioning coordinates and other information obtained after detection and segmentation, the width is measured, and the road defect passability risk level is identified and determined.

[0042] In summary, the algorithm reduces the computational complexity while maintaining high segmentation accuracy, improves the inference speed, can stably identify and track the target, and outputs the robot passability risk level information, so as to support the adaptive speed regulation of the robot, meet the lightweight deployment requirements of the mobile platform, and has important role in the fields of industry, security inspection, infrastructure maintenance and the like. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is the technical route of the passability determination algorithm of the wheeled robot under road defect conditions.

[0044] Figure 2 is a hardware implementation framework of a wheeled robot.

[0045] Figure 3 is a network structure diagram of a YOLO11n-RVD target detection segmentation model.

[0046] Figure 4 is a network structure diagram of a dual information capture module C2DIC.

[0047] Figure 5 is a network structure diagram of a designed SEPOKM structure.

[0048] Figure 6 is a network structure diagram of an SGCLDH lightweight detection head.

[0049] Figure 7 is a network structure diagram of a target enhanced convolution SGConv.

[0050] Figure 8 is a curve diagram of the precision of a YOLO11n_seg (A), YOLO11n-RVD (A_starnet-C2DIC-EP-SGCLDH) target detection segmentation model changing with training rounds; wherein, (a) precision; (b) recall; (c) average precision mean mAP_0.5; (d) average precision mean mAP_0.5:0.95.

[0051] Figure 9 is a curve diagram of the precision of three models YOLO11n_seg (A), YOLO11n_P2 (A_P2), YOLO11n_EP (A_EP) changing with training rounds; wherein, (a) precision; (b) recall; (c) average precision mean mAP_0.5; (d) average precision mean mAP_0.5:0.95.

[0052] Figure 10 is a comparison diagram of the results of three models YOLO11n_seg (A), YOLO11n_P2 (A_P2), YOLO11n_EP (A_EP) detecting road defect images; wherein, (a) original image; (b) A model detection; (c) A_P2 model detection; (d) A_EP model detection.

[0053] Figure 11 is a curve diagram of the precision of three models YOLO11n_seg (A), YOLO11n_CLDH (A_CLDH), YOLO11n_SGCLDH (A_SGCLDH) changing with training rounds; wherein, (a) precision; (b) recall; (c) average precision mean mAP_0.5; (d) average precision mean mAP_0.5:0.95.

[0054] Figure 12 are the scatter plots of total time consumption and accuracy of four model inferences YOLO11n_seg(A), YOLO11n_CLDH(A_CLDH), YOLO11n_SGCLDH(A_SGCLDH), YOLO11n-RVD(A_starnet-C2DIC-EP-SGCLDH).

[0055] Figure 13 are the result example graphs of real-time tracking detection of three defects; wherein (a) low risk; (b) medium risk; (c) high risk. DETAILED DESCRIPTION

[0056] The related technical solutions will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] A wheeled robot passability judgment algorithm under road defect conditions, the technical route is as shown in Figure 1 , comprising the following steps:

[0058] Step S1: design of the hardware system of the wheeled robot, including using radar and a depth camera for navigation, using a camera on the head of the wheeled robot to take images, obtaining a video stream of the inspection scene, transmitting it to a server, detecting the video stream by a target detection algorithm to obtain target classification information and positioning information, and tracking, the system implementation architecture is as shown in Figure 2 .

[0059] Step S2: taking pictures of various road defects and labeling to make labels, making a data set, specifically, based on the public data set Crack Segmentation Dataset, selecting the CFD and Crack500 sub-datasets related to road defects, removing non-defect images and wall crack samples in non-road scene, and retaining concrete road and asphalt road cracks. Then, using the Intel RealSense D435 depth camera fixedly installed on the head of the robot, 500 road defect images are taken at a walkable speed of 0-3.6 km / h, the original images collected are subjected to Gaussian blur and random noise data enhancement processing, the number of samples is expanded to 2000, and the labelimg tool is used to label the segmentation labels. Finally, the samples are unified, not less than 5000, divided into training set, verification set and test set according to the ratio of 7:1:2, and arranged in the standard directory structure.

[0060] Step S3: develop a YOLO11n-RVD target detection segmentation model to identify images or videos captured by a robot camera. The development of the YOLO11n-RVD target detection segmentation model includes the following steps:

[0061] Step S3.1: improve the backcone of the model using the starnet structure, because the global attention mechanism in the C2PSA module has high computational complexity and insufficient local detail capture, a two-stage spatial attention module TSSA is designed for feature extraction to replace the C2PSA module, which can better capture local details and global context information in the image while reducing computational complexity.

[0062] Step S3.2: improve the YOLO11n_seg model using an efficient small target layer detection feature pyramid EP, and replace the Bottleneck in the starnet structure with Star Blocks to form a C3k2_Star module. Through depth separable convolution and gating mechanism, the feature extraction capability is improved while reducing the parameter amount, which is especially suitable for mobile deployment. Its double-layer large core design expands the receptive field, and the residual structure ensures the training stability, finally realizes the double improvement of precision and efficiency.

[0063] Step S3.3: design a target strengthening convolution SGConv to improve the lightweight detection head, and propose a SGCLDH detection head.

[0064] Step S3.4: train the image dataset using the YOLO11n-RVD target detection segmentation model.

[0065] In step S3, a deep learning algorithm YOLO11n_seg is used to train and establish a target detection segmentation model. The target detection segmentation model includes an input end, a backbone network, a neck module, a head module, and an output end. The preprocessed image is input into the backbone network through the input end. The backbone network is used to extract features from the preprocessed image to extract feature maps with different feature scales. The neck module re-fuses the feature maps with different feature scales extracted by the backbone network into feature maps with different scales through upsampling and downsampling, and inputs them into the head module. The head module is used for target frame selection and target detection segmentation to obtain results. The present application develops a YOLO11n-RVD target detection segmentation model, which introduces a starnet structure and a C2DIC shared convolution network in the backbone network, introduces an EP small target detection layer and a Star Blocks in the Neck module, and introduces a SGCLDH target strengthening lightweight detection head in the Head module. The structure of the target detection segmentation model is as shown in Figure 3 .

[0066] In the step S3, the starnet_s050 in the starnet series is used to replace the backbone, which is the lightest backbone network variant in the starnet series, suitable for edge computing devices and video stream analysis. The starnet architecture maps low-dimensional input to high-dimensional nonlinear feature space through element-wise multiplication, reduces the parameter amount, captures long-distance dependence through depth separable convolution DWConv, and realizes efficient feature extraction through lightweight attention mechanism. The process from input to output is as follows, a total of four stages (k takes 1~4):

[0067]

[0068] A Dual Information Capture (C2DIC) module is developed to replace the C2PSA module. The original PSA Block uses the basic attention module Attention, while the improved TSSA Block uses the AttentionTSSA attention mechanism. This module adopts two-stage calculation (local + global attention): through grouping strategy to reduce the calculation complexity, from O(H²W²) to O(HW(M + G²)), where H and W are the height and width of the feature map, M is the local window size, and G is the number of groups. The local attention captures fine-grained details, and the global attention models long-distance dependence. The feature complementation is realized through addition operation:

[0069]

[0070]

[0071] In the formula, Q, K, and M are query matrix, key matrix, and value matrix, respectively; is a scaling factor; is the dot product of the query and the key; is the linear transformation matrix after the multi-head attention output, which is used to splice and map the multi-head results back to the target dimension; is the output of the i-th attention head, i is 1, 2, …, h, and h is the number of attention heads. The improved module can better capture the long-term spatial dependence of the feature map and improve the detection accuracy of small or occluded objects. Its network structure is shown in Figure 4

[0072] ​In the step S3, small-scale targets are not easy to be detected because small target features are lost in the downsampling process. To improve the detection capability, the traditional method is to add a P2 detection layer, which provides higher spatial resolution and retains more small target details, but also causes the problem of excessive calculation after adding P2 and more time-consuming post-processing. The present application is improved based on PAFPN, and an Efficient Small Object Pyramid (EP multi-scale detection network) is developed, so that the P2 feature layer is subjected to slicing operation by SPDConv to obtain features rich in small target information for P3 fusion:

[0073]

[0074]

[0075] In the formula, is the input feature map from the P2 feature layer, is the input feature map from the P3 feature layer, represents a feature splicing operation, is the output feature map of the enhanced P3 feature layer after fusion.

[0076] Based on the CSP structure, the input features are divided into two parts, the OKM branch and the Initial original feature branch. Due to the large size of the Omni-Kernel module, the branch is set to occupy 25% of the channel number, combined with multi-scale convolution and frequency domain enhancement, to realize efficient feature integration, as shown in Figure 5 The FGM module is a frequency-guided modulation module that fuses spatial and frequency domain information to enhance feature expression. The OKM is a multi-scale convolution and frequency space attention mechanism that combines multi-scale convolution, frequency channel attention FCA, spatial channel attention SCA and FGM to extract multi-scale spatial features and global frequency features. SEPOKM is used to integrate OKM and Initial branch, and through its global branch, large branch and local branch, the feature representation from global to local is effectively learned to finally improve the detection performance of small targets:

[0077]

[0078]

[0079]

[0080] In the formula, the feature map where B is batch size, C is the number of channels, HxW is spatial resolution, FFT is two-dimensional fast Fourier transform, IFFT is inverse Fourier transform, are learnable parameters, the fused feature map .

[0081] In the step S3, in order to make the model more lightweight, reduce the number of parameters, and speed up the inference, the application proposes a Strengthening Convolutional Lightweight Detection Head, namely SGCLDH detection head, which uses a target strengthening convolution SGConv as an intermediate shared convolution for sharing feature enhancement, and the SGCLDH detection head structure is as shown in Figure 6 The SGConv adopts a multi-branch convolution design, fuses a center difference Conv2d_cd, a horizontal difference Conv2d_hd, a vertical difference Conv2d_vd, an angle difference Conv2d_ad, and a standard convolution, generates an equivalent 3x3 convolution kernel through weight fusion, enhances the details of the features, is suitable for high-precision detection tasks, and the SGConv module is as shown in Figure 7 The implementation details of the SGConv module are as follows:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] In the formula, the feature map , is a weight matrix of each branch convolution, is a bias term of each branch convolution, and GroupNorm 16 is normalized by 16 groups on the channel, which can improve the positioning and classification functions of the detection head. In order to cope with the problem that the target scales detected by each detection head are inconsistent, a Scale layer is used to scale the features, reducing the loss of precision. The SGConv integrates prior information into the ordinary convolution layer to enhance the representation and generalization ability, and then converts the SGConv into an ordinary convolution through the use of the reparameterization technique, so that fast inference can be realized without additional parameters and calculation costs.

[0088] Step S4: using a tracking algorithm to locate the recognition result in the image, and outputting a risk value combined with a risk determination algorithm.

[0089] In the step S4, the image is captured in real time by the camera carried by the robot, the pixel of the image is 1920*1080, the image or video stream is input into the YOLO11n-RVD model for recognition, the BoT-SORT algorithm is used to track multiple targets in the image, and target classification and positioning information are obtained.

[0090] The present application adopts a pavement defect-tire width ratio L as a core index for evaluating the static risk of robot passability, and the key risk source is whether the tire can lose support and fall into the defect because the defect is too wide, and the axial width of the tire determines its ability to cross the seam. The ground contact size in the rolling direction of the tire mainly affects the dynamic vibration and impact load when passing through the defect, which is a secondary factor for the low-speed inspection robot in the present application. Therefore, the present application focuses on the falling risk defined by the static ground contact width, and this index can more directly and simply serve the decision goal of passability safety warning. The pavement defect-tire width ratio L is calculated as follows:

[0091]

[0092] In the formula, is the width of the widest part of the defect, is the tire ground contact width under static. Based on this method, the present application establishes a three-level evaluation threshold system of pavement defect passability risk as shown in Table 1, and further couples the risk level with the motion control strategy of the robot: maintaining the original speed under low-risk conditions; introducing a speed feedback mechanism for moderate deceleration to pass through cautiously under medium-risk conditions; triggering braking behavior immediately to achieve emergency parking and avoid falling under high-risk conditions. In order to accurately obtain the physical width of the pavement defect, the camera is first calibrated to obtain its intrinsic and extrinsic matrices and distortion coefficients. Then, combined with the installation height and pitch angle of the camera relative to the road, a monocular vision geometric measurement model is constructed, and the pixel width of the defect in the image is converted into the actual physical width through the principle of triangulation. Finally, the system outputs the corresponding detection and risk assessment results according to the calculated physical width and its risk level, and feeds back to the motion control unit in real time to realize adaptive speed regulation.

[0093] Table 1 Risk evaluation threshold

[0094]

[0095] In order to improve the practicability and interactivity of the system, and increase the observability, the application develops a graphical interface based on python for real-time monitoring. The graphical interface is based on PyQt5 module, integrates YOLO model and OpenCV vision processing library, supports dynamic model loading, real-time video / camera input, detection and tracking functions, and provides pause, frame saving and video recording operation options. Users can specify defects for real-time tracking through the target ID search box. The system interface displays the defect trajectory and its bounding box center pixel coordinates in real time, visualizes the tracking process, and outputs the risk determination results: risk level, risk description L, width of the widest part of the defect, distance of the defect bounding box center from the camera, confidence and angle of the defect extension direction, and real-time feedback of the current speed control state.

[0096] The application is further explained in conjunction with the following drawings, examples and comparative examples:

[0097] Examples

[0098] A kind of road defect condition under wheeled robot passability determination algorithm, this method is realized using the following steps:

[0099] Step S1: a self-built wheeled robot is used, and a camera mounted thereon is used for image acquisition.

[0100] Step S2: during the movement of the robot at a speed of 0-3.6 km / h, the camera parameters including exposure time, ISO sensitivity and white balance are adjusted according to the road defect environment on various roads to adapt to dynamic light changes. According to the above settings, multi-angle shooting is used to simulate the actual inspection viewing angle, enhancing the perspective diversity and dynamic reality of the data set. Finally, the obtained image data is divided into training set, validation set and test set in the ratio of 7:1:2, totaling 10000, and unified to the standard directory structure.

[0101] Step S3: In a Windows 10 system environment, the hardware configuration is Intel® Xeon(R) Platinum 8375C CPU, NVIDIA RTX A6000. The deep learning framework uses PyTorch 2.4.1, and cooperates with CUDA 11.8. The training setting is 500 epochs, the batch size is 16, the optimization algorithm is stochastic gradient descent (SGD), and the resolution of the input image is 640x640. The purpose of the present application is to realize model lightweight while maintaining high accuracy of monitoring. Therefore, the precision, recall, mean average precision (mAP_0.5), mAP_0.5:0.95, model parameter quantity (Parameters / M), gigaflops per second (GFLOPs), and frames per second (FPS) of the model are selected as evaluation indexes. AP is the area under the precision-recall curve. i The prepared data set is input into the YOLO11n-RVD target detection segmentation model for training, the training configuration is the same, and the trained different models are obtained. The best.pt file is obtained by training the training set, and the important parameters of the model are obtained by using the validation set, thereby obtaining the YOLO11n-RVD target detection segmentation model.

[0102] Step S4: The BoT-SORT algorithm is used for multi-target tracking of the recognition results in the image, the defect width is measured by triangulation, and the pass risk is determined.

[0103] To verify the effectiveness of each module in the YOLO11n-RVD model, the following ablation experiments are performed:

[0104] Ablation experiment 1:

[0105] Steps 1-4 of the experiment are the same as steps S1-S4 of the embodiment. In step 3 of the experiment, the original YOLO11n_seg model is used to train the target detection segmentation of the self-built data set.

[0106] Ablation experiment 2:

[0107] Steps 1-4 of the experiment are the same as steps S1-S4 of the embodiment. In step 3 of the experiment, the YOLO11n_seg model is used, and the starnet star-shaped structure and C2DIC double information capture structure are introduced to improve the backbone network, thereby obtaining the target detection segmentation model, and training the self-built data set.

[0108] Ablation experiment 3:

[0109] The experimental steps 1-4 are the same as the example steps S1-S4, in this experiment, the original model is introduced into the small target detection layer EP structure and the Star Blocks improved neck module in step 3, and the self-built data set is trained;

[0110] Ablation experiment 4:

[0111] The experimental steps 1-4 are the same as the example steps S1-S4, in this experiment, the original model is introduced into the small target detection layer EP structure and the Star Blocks improved neck module in step 3, and the self-built data set is trained.

[0112] Ablation experiment 5:

[0113] The experimental steps 1-4 are the same as the example steps S1-S4, in this experiment, the original model is introduced into the small target detection layer EP structure and the Star Blocks improved neck module in step 3, and the self-built data set is trained.

[0114] The ablation experiment results are shown in Table 2.

[0115] Table 2 Comparison of ablation experiment results

[0116]

[0117] From the ablation experiment results, after introducing the starnet-C2DIC structure of the application, the parameter quantity is reduced from 2.83M to 2.18M, the calculation amount FLOPs is reduced from 5.8G to 4.5G, the model size is reduced from 5.8MB to 4.5MB, and the model inference speed is increased by 13.2%, indicating that the starnet-C2DIC network can effectively reduce the model complexity and stabilize the detection performance. From experiments 1 and 3, the improvement of the neck layer greatly improves the performance, and mAP_0.5 increases by 2.9%. From experiments 1 and 4, the introduction of the target strengthening lightweight detection head SGCLDH structure not only reduces the complexity of the model, but also increases the model inference speed by 41.0%, and mAP_0.5 is improved by 1.5%, which undoubtedly improves the network performance.

[0118] Combining the three improvements, according to the experiment 5, the YOLO11n-RVD model combines the advantages of small target improvement and lightweight, the parameter quantity is further reduced to 1.92M, the calculation quantity is reduced from 16.2 to 10.9 GFLOPs in the Neck layer, the model size is reduced to 4.6MB, and the mAP_0.5 is increased by 3.1%, the mAP_0.5:0.95 is increased by 1.8%, and the inference speed is increased by 25.2%, which shows that the combination of experiment 5 realizes the improvement of precision and achieves the lightweight target, and the precision comparison curve is as shown in Figure 8 .

[0119] Comparative Example 1

[0120] The small target detection layer P2 structure is introduced to improve the Neck module, compared with the ablation experiment 3 of the present application. The comparison results are shown in Table 3.

[0121] Table 3 Comparison of small target detection improvement model experiment results

[0122]

[0123] Comparing the experimental results, compared with the P2 layer improvement, the EP module of the present application increases the parameter quantity by 13.4%, but reduces the calculation quantity by 33.5%, increases the mAP_0.5 by 0.8%, and also increases the mAP _ 0.5:0.95 by 0.4%, and the model inference time is 10.6% faster, and the precision comparison curve is as shown in Figure 9 This shows that EP can improve the detection performance of small targets and speed up the processing speed. For multi-scale road defects, as shown in Figure 10 (a), the detection results of the original model, the improved P2 model and the improved EP model are shown in Figure 10 (b), (c) and (d), respectively, and the analysis shows that the small target detection improvement is effective, and the EP model is better than the P2 improved model.

[0124] Comparative Example 2

[0125] The lightweight detection head CLDH structure is introduced to improve the Head module, compared with the ablation experiment 4 of the present application. The comparison results are shown in Table 4.

[0126] Table 4 Comparison of detection head improvement model experiment results

[0127]

[0128] Comparing the experimental results, the SGCLDH module improves the mAP_0.5 by only 0.3% compared with the CLDH improvement, and the precision comparison curve is as shown in Figure 11The parameter quantity is reduced by 6.3%, and the inference speed is increased by 37.6%. The total time consumption (preprocessing time plus inference and post-processing time) of the four models YOLO11n_seg, YOLO11n-CLDH, YOLO11n-SGCLDH and YOLO11n-LS for inferring the verification set pictures is taken as the horizontal coordinate, and the accuracy of each model is taken as the vertical coordinate, as shown in Table 5. Figure 12 As shown in Table 5, the SGCLDH module proves to be effective in improving the inference detection speed.

[0129] Comparative Example 3

[0130] A plurality of lightweight target detection models of the YOLO series widely used at present and with open codes are compared and analyzed under the condition of unified experimental settings and data sets. The selected comparative models are all derived from different versions of the YOLO series, and the lightest models in the selected versions are selected, which covers representative methods of different complexities and structures and can comprehensively reflect the performance differences of lightweight algorithms in the road defect detection task. All models are trained and tested on the self-built image data set covering multiple typical road scenes to ensure the fairness of the comparison and the reliability of the results. The detailed comparison results are shown in Table 5. The experimental results are shown in Table 5.

[0131] Table 5 Comparison of different models

[0132]

[0133] As can be seen from the analysis, the baseline model YOLO11n_seg selected by the present application has similar accuracy to YOLOv9t_seg on the self-built data set and is higher than other basic models, but the computational load of the baseline model is less than that of the YOLOv9t_seg model, the model size is slightly higher than that of YOLOv10n_seg, and the inference speed is lower than that of YOLOv8n_seg. After improvement, YOLO11n-RVD has the highest defect detection and segmentation accuracy, the lowest network size, and only slightly lower inference speed than the YOLOv8n_seg model, which well achieves the research goal of the present application, i.e., improving the accuracy of the model, reducing the structural complexity, and improving the inference speed.

[0134] As can be seen from the combination of the embodiments and the comparative examples, the YOLO11n_seg target detection and segmentation model realizes significant optimization of model complexity and computational cost by introducing the starnet and SGCLDH components, realizes high-precision detection of multi-scale targets by introducing the EP layer, and speeds up the inference speed by introducing the C2DIC and SGCLDH modules. Therefore, the YOLO11n-RVD model effectively realizes the balance between detection performance and lightweight.

[0135] In summary, the road defect monitoring algorithm combining image segmentation and width measurement proposed by the present application can realize the positioning of defects and the automatic extraction of morphological parameters (including but not limited to width, length and direction). On this basis, combined with the physical size of the robot tire, by calculating the ratio of defect width to tire ground width, a three-level early warning mechanism of "low risk", "medium risk" and "high risk" is set, which provides quantitative basis for real-time path planning and safety warning of wheeled robots. At the same time, the present application focuses on the high-precision detection and segmentation of road defects and the width measurement algorithm, and integrates the robot passability judgment framework, which provides a reliable visual perception solution for the autonomous inspection and safe passage of wheeled robots in complex road environments.

[0136] In the implementation process, the present application adopts the static ground width of the tire The real-time tracking results of defects are shown in Figs. Figure 13 (a), (b) and (c), which show examples of tracking two targets, Figure 13 In (a), the defect is low risk, and the following indicators are output: confidence 0.69, L=0.20, width 5.1mm, distance 267.5mm, angle 7.5°, maintain current speed; Figure 13 In (b), the defect is medium risk, and the following indicators are output: confidence 0.66, L=0.35, width 8.6mm, distance 298.9mm, angle 35.6°, trigger deceleration mechanism, pass carefully; Figure 13 In (c), the defect is high risk, and the following indicators are output: confidence 0.83, L=1.35, width 33.7mm, distance 526.1mm, angle 31.6°, emergency brake, reverse turn and re-plan path.

[0137] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An algorithm for determining the passability of a wheeled robot under road defect conditions, characterized in that: Includes the following steps: S1: Design a wheeled robot hardware system, including a camera for real-time acquisition of road surface images; S2: Construct an image dataset containing various road surface defects and annotate it; S3: Based on the YOLO11n-seg model, develop the YOLO11n-RVD target detection and segmentation model to identify and segment road surface defects in input images; wherein, the YOLO11n-RVD target detection and segmentation model introduces a StarNet structure and a C2DIC module in the backbone network, introduces an EP small target detection layer and Star Blocks in the neck network, and introduces an SGCLDH lightweight detection head in the head network; S4: Use a multi-target tracking algorithm to track the identified defects and combine it with a risk assessment algorithm to output the passability risk level.

2. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S3, the backbone network adopts the Starnet_S050 structure and replaces the original C2PSA module with the C2DIC module. The C2DIC module adopts a two-stage spatial attention TSSA mechanism. First, local attention is calculated within a local window to capture fine-grained details. Then, the feature maps are grouped using a grouping strategy, and global attention is calculated within each group to model long-distance dependencies. Finally, the outputs of local attention and global attention are added to achieve feature complementarity. The grouping strategy reduces the space complexity of attention calculation from O(H²W²) to O(HW(M + G²)) by grouping the channel dimensions of the key K and value V, where H and W are the height and width of the feature map, M is the size of the local window, and G is the number of groups.

3. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S3, the EP small target detection layer is based on an improved PAFPN structure. It enhances the preservation and extraction of small target features by performing an SPDConv slicing operation on the P2 feature layer and then fusing it with the P3 feature layer. The slicing operation is implemented using the following formula: ； ； In the formula, The input feature map comes from the P2 feature layer. The input feature map comes from the P3 feature layer. This indicates a feature concatenation operation. The output feature map is the enhanced P3 feature layer after fusion.

4. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 3, characterized in that: The EP small target detection layer includes a SEPOKM structure, which integrates the OKM branch and the Initial branch, and integrates a frequency-guided modulation module and a multi-scale convolution mechanism to improve the detection accuracy of multi-scale defects.

5. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S3, the SGCLDH lightweight detection head uses the SGConv module. This module generates an equivalent convolution kernel by fusing the weights and biases of multi-branch convolutions to achieve efficient feature enhancement. The process is defined by the following formula: ； ； ； In the formula, , These are the weight matrices and bias terms for the center-discrimination convolution, horizontal-discrimination convolution, vertical-discrimination convolution, angular-discrimination convolution, and standard convolution, respectively. For the input feature map, For convolution operations, The weight matrix after fusion. For the fused bias term, This is the output feature map after convolution calculation.

6. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 5, characterized in that: The output of the SGConv module is processed sequentially by group normalization and the SiLU activation function: ； ； In the formula, GroupNorm 16 This indicates that the channels are normalized according to 16 groups.

7. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S4, the BoT-SORT algorithm is used to perform multi-target tracking on the identified defects to obtain the location, ID and motion trajectory information of the defects.

8. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S4, the risk assessment algorithm calculates the width of the widest point of the defect. Tire contact width under static conditions The ratio L, i.e., the ratio of road surface defect to tire width, is used to set three levels of risk thresholds.

9. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: In step S2, the image dataset contains no less than 5,000 images of road surface defects, which are divided into training set, validation set and test set in a ratio of 7:1:

2.

10. The algorithm for determining the passability of a wheeled robot under road defect conditions according to claim 1, characterized in that: The wheeled robot passability determination algorithm also includes a graphical interface, developed based on PyQt5, for real-time display of defect detection results, tracking trajectory, risk level and robot motion control status, and supports user interaction and data recording.

Images