Deep Learning Model for Pothole Detection and Area Computation

Pothole is the major type of structural defect found on road commonly generated due to structure aging, large rainfall, dense traffic and thin or weak substructure, etc. Day by day regular maintenance and assessment of such roads becoming a very challenging task. In this paper, a deep learning technique Mask Region-Based Convolutional Neural Network is proposed to accurately detect and segment such potholes to predict and calculate its area. The database of 291 images used which has been collected manually on local roads of Mumbai city and nearby highways. The dataset is annotated using the freely available VGG Image Annotator manual tool. Using Mask Region-Based Convolutional Neural Network (Mask RCNN) potholes are detected as a region of interest. And based on this generated region of interest area of the pothole is computed. The computed pothole area is then compared with the actual measured area. The proposed methodology presented in this paper has been implemented in python under a Windows and colab environment. The proposed system shows that the region of interest from the pothole images is quite well detected. Experimental results give an overall accuracy of 90% for the computed area of the pothole when compared with the actual measured area with ± 10% deviation. Results are promising, and the information extracted using the proposed method can be used for cost estimation in road repair management.