| Transportation industry is the foundation of the rapid and sustainable development of the national economy.With the enhancement of China’s strength,great process has been made in the development of highway traffic.Compared with the developed countries,China’s highway construction started late,and the asphalt pavement is likely to have the problems such as early damage and insufficient durability.Besides management and construction,insufficient focus on gradation and morphological characteristics of aggregates are also a very importance factor.Optical image measurement method can realize nondestructive and fast automatic detection of mineral mixture quality.However,most of aggregate characteristic image detection methods are based on two-dimensional particle projection,and its limitation of three-dimensional(3D)characterization can result in the limited accuracy.In the field of aggregate morphology characterization,the majority of detection methods are aggregate sampling and then testing in laboratory.Therefore,this dissertation studies the key methods and techniques of image filtering,aggregate particle segmentation,particle sieve-size classification and angularity characterization for three-dimensional aggregate particle image.Laser triangulation and structured light 3D camera are used to construct aggregate particle morphology acquisition system,which realizes the acquisition and data processing of aggregate particle 3D point cloud data and depth data.The background direct subtraction technique is used to solve the problem that the measured aggregate height deviates from the actual height.This method restores the actual height of aggregate.The Depth-Binary algorithm based on alternating sequence filter(ASF)is proposed to solve the noise removing problem in the depth image after background subtraction,which can smooth the depth images and keep the actual height value of aggregates.The deep watershed algorithm based on edge probability image and the nearest neighbor methods are proposed to segment the piled and separated aggregate particles respectively.In the proposed deep watershed algorithm,the grayscale depth images are used as input images of the edge detection model,and its output edge probability images are used as the energy graph in the watershed algorithm.At the same time,the markers are extracted from the distance transform of the thinned edge probability images,and then the aggregates are segmented using marked watershed algorithm.The results demonstrate that this method can effectively suppress the over-segmentation problem in watershed algorithm.The nearest neighbor 3D particle segmentation method takes the 3D aggregate point cloud data as the segmentation object.The adjacent distance of each point in the point cloud in the two-dimensional projection region is taken as the point merging criterion.This method has the advantage of end-to-end segmentation for the separated 3D point aggregate particles.Because of the occlusion problem in laser triangle vision system,the data missing can be easily induced in the collected 3D aggregate particles,which has a big influence on the detection accuracy of the vision system.In this dissertation,a particle contour detection method with edge fusion is proposed.The method combines the data characteristics of the occluded area and extracts its edge information and then integrates the extracted edge information into the contour information of the particle,which can effectively complete the lost edge information of the detected particle.Based on the fused edge information,the marked watershed algorithm is used to segment the particles.The experimental results show that the method is effective in detecting single-layer particles.To tackle the problem of aggregate sieve-size classification cannot be effectively realized by the traditional single particle size descriptors,the supervised machine learning methods for particle sieve-size classification prediction are used.In this dissertation,a dataset of 2D/3D features of aggregate particles is created.A feature selection algorithm with feature importance score priority was proposed and evaluated on 17 different machine learning classification algorithms.The experimental results show that the Gaussian Processes Classifier(GPC)achieves the highest classification performance in different subsets.The accuracy of the 3D dataset was up to 95.06% and was much higher than that of the single particle size characterization descriptor.To tackle the problem of angularity evaluation of piled aggregate particles,a new method called the Virtual Cutting Method is proposed.In this method,a virtual 3D plane is used to cut the upper surface of 3D aggregate object to obtain the 2D intersection lines based which a gradient method is used to calculate the angularity index of aggregates.Scatter plot shape,execution time and single-factor variance results of different size and texture particles at different cutting angles were analyzed.The experimental results show that the cutting Angle of 5° can achieve a balance between performance and execution time.In addition,the angularity index results of the Virtual Cutting Method were compared with those of 2D and 3D Projection Methods,experimental results show that the Virtual Cutting Method is more accurate and robust in the characterization of aggregate angularity,which can be used to evaluate the angularity index of piled aggregates dynamically.Based on 3D point cloud data of aggregate particles,this dissertation proposes 3D aggregate segmentation,particle edge detection under camera occlusion,aggregate sieve-size classification and aggregate angularity evaluation using the Virtual Cutting Method,effectively solving some shortcomings of existing methods.It will establish theoretic and technical foundation for the intelligent quality detection of aggregate particles. |
