Home Page
» Online Library » Recognition and Clustering of Road Pavement Defects by Deep Machine Learning Methods
Recognition and Clustering of Road Pavement Defects by Deep Machine Learning Methods
| Theme: |
Downloading for authorized users
|
|
| Author: | Finogeev Alexey , Финогеев Антон Алексеевич , (with the team of authors) | |
| Annotation: | The paper discusses an approach to detecting and classifying pavement defects in images of road sections obtained from photographs and video frames in the process of road scanning. A modern way to reduce the risks of road incidents is to implement technologies for monitoring and analyzing various factors in order to assess their impact on accident rates. One of the factors is pavement defects, which deteriorate the performance of roads and become the causes of road incidents and vehicle damage. The aim of the study is to recognize roadway defects based on their degree of contrast compared to defect-free sections. The main problems are the accuracy of recognizing and classifying defects under conditions of noise associated with weather conditions and lighting, as well as the inaccuracy of determining their geometric dimensions from two-dimensional images. To solve the problems of detection, recognition and classification of pavement defects with acceptable accuracy in image processing with noise, we developed an approach that includes two methods: a) a method using IoU-HOG-ACM-BoVW algorithms for recognition and classification of high noise images, b) a method of segmentation and recognition using convolutional neural network MaskR-CNN for recognition and classification of low noise images. The algorithms for solving the problems of segmenting road pavement images into parts with detected damages, determining the boundaries and geometric dimensions of defects, classifying damages by type, clustering and ranking road segments according to the operational condition of the road pavement to reduce accidents, improve traffic safety and planning repair work on highways are considered. The proposed approach includes two methods: the first is a method for estimating road segments based on their clustering using the DBSCAN algorithm, taking into account noise and the distribution density of similar segments on a digital map. The method identifies clusters of dangerous road segments and ranks them according to the operational condition of the road surface. The approach increases the level of road safety and allows efficient allocation of resources for repair and preventive works. The paper discusses a methodology for assessing road safety in selected clusters of road sections. A noise-aware clustering algorithm can identify the distribution densities of similar sections with different defects, identify patterns of their occurrence, detect anomalies, and provide decision-making for repair planning. Imagine a scenario where accidents and maintenance costs can be reduced by identifying problem road sections in advance | |
| Type: | Article | |
| Kind: | Electronic copy | |
| Parts: | 1 | |
| The year of publishing: | 2024 | |
| Publishing house: | Machine Learning Methods in Systems. CSOC 2024. | |
| Publishing place: | Machine Learning Methods in Systems Proceedings of 13th Computer Science On-line Conference 2024, Vol. 4 | |
| The target audience: | Researcher | |
| Special purpose: | Scientific | |
| Copyright holder: | Финогеев А.А., Финогеев А.Г., Деев М.В., Парыгин Д.С., При поддержке Российского Научного Фонда | |
| ISBN: | 978-3-031-70595-3 | |
| DOI: | https://doi.org/10.1007/978-3-031-70595-3_48 | |
| Bibliographic reference: | Finogeev, A., Deev, M., Finogeev, A., Parygin, D. (2024). Recognition and Clustering of Road Pavement Defects by Deep Machine Learning Methods. In: Silhavy, R., Silhavy, P. (eds) Machine Learning Methods in Systems. CSOC 2024. Lecture Notes in Networks and Systems, vol 1126. Springer, Cham. https://doi.org/10.1007/978-3-031-70595-3_48 | |
| Url: | https://link.springer.com/chapter/10.1007/978-3-031-70595-3_48#citeas | |
| Language: | English | |
| Post date: | 28.10.2024 |