Multi-scale Texture Feature Extraction And Skid Resistance Performance Evaluation Of Asphalt Pavement Based On Point Cloud Data

Asphalt pavement skid resistance is one of the important factors affecting road safety,with good skid resistance performance of the pavement can provide sufficient friction for the vehicle,thereby reducing the probability of accidents.The surface texture structure is a direct factor affecting the skid resistance of pavement and determines the contact state between vehicle tires and the road surface.Therefore,it is important to study the influence of pavement texture on skid resistance performance to improve road safety performance.A single index cannot fully characterize the pavement texture details and reveal the mechanism of pavement skid resistance evolution.In this paper,asphalt pavement 3D texture point cloud data are acquired based on binocular stereo vision system.Then,the intrinsic correlation mechanism between multi-scale texture characteristics and asphalt pavement skid resistance is studied,and an intelligent evaluation model of skid resistance under different scenarios is established.The main research work and innovations carried out are as follows:(1)A method for abnormal diagnosis and quality control of 3D texture point cloud data is proposed.To solve the problem of missing values in 3D texture point cloud data of road surface collected by precision optical equipment,a calculation method of missing values generated by adversarial inference network GAIN is proposed.Compared with the original GAN network,the GAIN network structure adds a prompt mechanism to continuously strengthen the antagonism process between the generator and discriminator,thus improving the filling accuracy of missing values.Compared with the classical KNN algorithm,the accuracy of filling missing data of SMA-16 and AC-16 in the GAIN network is improved by 4.8% and 5.8%,respectively.To accurately identify and locate outliers in 3D texture point cloud data,an outlier detection method based on unsupervised fusion model DBSCAN-i Forest is proposed.Firstly,DBSCAN is used to cluster the original 3D point cloud data.Then,isolated forest i Forest algorithm is used to identify outliers in discrete categories.The outlier detection results of the method are evaluated by Davidson-Boding index DBI.The smaller the DBI value is,the better the outlier detection algorithm performance is.Compared with the classical COF detection algorithm,the DBI value of the unsupervised fusion model DBSCAN-i Forest is the smallest,and the DBI value on the three-dimensional texture point cloud dataset of AC-16 is 0.3001.(2)A multi-scale texture feature extraction and correlation analysis method for 3D point cloud data is designed.To eliminate the boundary effect of traditional Gaussian filter algorithm,a macro-micro texture separator based on robust Gaussian filter is designed according to the wavelength range and separation criteria of macro-micro texture.At the same time,height,function,space,composite,and volume parameters are extracted from macro and micro texture data to characterize the three-dimensional texture information of asphalt pavement.To study the correlation between texture features and skid resistance performance at different scales,the two-dimensional gray image is decomposed into four different scales based on the discrete wavelet transform algorithm.For each scale,the Angle second moment,entropy,contrast and correlation indexes are extracted by gray-level co-occurrence matrix to describe the characteristics of multi-scale macro texture images of asphalt pavement.Aiming at the interactive influence problem among the constructed asphalt pavement texture features,correlation coefficients among macro texture features,micro texture features and multi-scale macro texture image features are calculated based on Pearson method,and redundant features with strong linear correlation are removed to reduce the interference between interaction features in the modeling process of pavement friction coefficient prediction.Improve the accuracy of pavement skid resistance performance prediction model.(3)An Optuna-NGBoost based asphalt pavement friction coefficient prediction model is proposed.To explore the complex nonlinear relationship between texture features and surface friction coefficient,a 3D texture feature importance automatic calculation model based on the fusion of gradient boosting algorithm is constructed.Multi-features of three-dimensional texture are input into the model.According to the results of feature importance ranking,macro and micro texture evaluation indexes with high correlation with pavement skid resistance performance and representative are selected.On this basis,to solve the problem of accurate evaluation of pavement skid resistance performance under small sample conditions,a prediction model of asphalt mixture surface friction coefficient based on Optuna-NGBoost is proposed.In this model,Optuna parameter optimization algorithm is adopted to find the optimal solution of NGBoost core parameters in the set global space domain,so as to adjust the model structure to the best.The Optuna-NGBoost fusion model is trained and verified with the surface friction coefficient as the target value.The results show that the Optuna-NGBoost fusion model can accurately predict the friction coefficient value with the prediction accuracy of 97.62%.Compared with XGBoost,Cat Boost and Light GBM models,the R~2 values of Optuna-NGBoost fusion model increased by 9.7%,8.6% and 0.7%,respectively.(4)A dynamic friction coefficient of asphalt pavement prediction method based on s GBM-Tab Net model is proposed.The accurate prediction of friction coefficient is a complex nonlinear problem due to the influence of several complex factors.For the constructed nonlinear and multi-scale asphalt pavement texture feature dataset,the conventional method has limited prediction accuracy.To solve this problem,on the basis of selecting texture features with high correlation and representing skid resistance of pavement,four heterogeneous gradient boosting models and Tab Net networks are fused based on stacking framework.Then,a dynamic friction coefficient prediction model based on s GBM-Tab Net fusion for asphalt pavement is constructed to enhance the prediction accuracy of the single model.The s GBM-Tab Net model is trained and verified with the dynamic friction coefficient as the target.The results show that the model can accurately predict the dynamic friction coefficient through multi-scale texture features and the prediction accuracy reaches 98.23%.Compared with MLP,Cat Boost,Light GBM,XGBoost,NGBoost and Tab Net models,the R~2 values of fusion model s GBM-Tab Net increased by 16%,5%,4%,3.8%,3% and 4.6%,respectively.In this research,a gradient boosting fusion model is proposed to reveal the internal correlation mechanism between texture multi-features and skid resistance performance.An intelligent prediction method for asphalt pavement skid resistance performance under small sample,multi-scale and multi-feature conditions is proposed,which solved the problems of sensitive parameter setting,poor network stability and nonlinear fitting performance of existing methods for processing such data,and improved the prediction accuracy of pavement skid resistance performance under complex data conditions.The research results provide theoretical basis and method support for the intelligent detection and evaluation of pavement skid resistance performance,and also have important significance for the realization of modern road safety performance information,digital,intelligent detection and management.