1348171 - A Single-Shot Model for Pavement Distress Detection and Quantification

Over the past decade, automated detection of pavement distress has remained a challenging task for computer vision researchers. However, recent advances in deep learning have stimulated a substantial amount of research aimed at improving the accuracy of automated pavement distress detection and classification. An efficient and effective distress evaluation model must 1) detect and classify pavement distress, 2) estimate the extent of the detected distress, and 3) estimate the severity of the detected distress concurrently. To date, there have been no existing datasets that provide all three pieces of information required for the efficient evaluation of pavement distresses. As a result, the current study aims to develop a novel benchmarked dataset that contains all three of the aforementioned pieces of information necessary for the efficient evaluation of pavement distresses. The dataset consists of pavement images captured from a top-down view. Each distress type was annotated manually with a bounding box and classified into one of nine distress categories. A mixture of manual and automated process such as adaptive thresholding techniques are subsequently used to extract distress patterns from the images to compute the extent and severity of each distress. Approximately 5,000 images were annotated. The new dataset was evaluated using a proposed multi-tasking deep learning framework capable of simultaneously detecting and classifying various distresses, estimating the extent of the distress, as well as estimating its severity. The proposed model inherits the architecture of the state-of-the-art object recognition system, YOLOv3, and extends its architecture by incorporating extent and severity, modifying its loss functions and prediction vectors to enable it to multitask on all four tasks. The model was built and implemented on top of the PyTorch framework. The proposed model was trained on an NVIDIA GTX 1080ti GPU. The new dataset and the benchmark models developed will drive a new generation of machine learning algorithms for the practical evaluation of pavement conditions.