Real-world Pothole Detection Using Image Processing and Deep Learning Convolutional Neural Network Model

Abstract

Potholes are a major problem of concern in many parts of the cities across the country. Road accidents are one of the causes that significantly affect humanity and result in damage to vehicles and road surface. Potholes are dangerous for pedestrians who walk along the road and vehicular traffic on busy roads. Road accidents are caused due to improper maintenance of roads, and it is imperative to attend to such hazards. Potholes should be maintained continuously to assure a minimal loss of lives and for the welfare of the community. The primary goal of this study is to build a convolutional neural network (CNN) model that identifies two classes by extracting region of interest (ROI) with minimum computational resource usage. This research introduces a deep learning approach termed CNN for the purposes of image segmentation and classification. Rectified linear unit (ReLU) activation function is used for classifying the input image and CNN model for feature extraction. After evaluation the accuracy we got is 87.77%. The approach discussed is relevant to concrete roads of all kinds, utilizing a dataset comprising 1157 images. In this process when a user uploads, the image of a pothole the proposed model will detect whether the image contains pothole or not. If the pothole is detected, then the image is converted to three-dimensional image to find the depth of a pothole above a threshold value. Finally, the information will be shared with government authorities to perform necessary action to ensure that no discrepancies occur.Potholes are a major problem of concern in many parts of the cities across the country. Road accidents are one of the causes that significantly affect humanity and result in damage to vehicles and road surface. Potholes are dangerous for pedestrians who walk along the road and vehicular traffic on busy roads. Road accidents are caused due to improper maintenance of roads, and it is imperative to attend to such hazards. Potholes should be maintained continuously to assure a minimal loss of lives and for the welfare of the community. The primary goal of this study is to build a convolutional neural network (CNN) model that identifies two classes by extracting region of interest (ROI) with minimum computational resource usage. This research introduces a deep learning approach termed CNN for the purposes of image segmentation and classification. Rectified linear unit (ReLU) activation function is used for classifying the input image and CNN model for feature extraction. After evaluation the accuracy we got is 87.77%. The approach discussed is relevant to concrete roads of all kinds, utilizing a dataset comprising 1157 images. In this process when a user uploads, the image of a pothole the proposed model will detect whether the image contains pothole or not. If the pothole is detected, then the image is converted to three-dimensional image to find the depth of a pothole above a threshold value. Finally, the information will be shared with government authorities to perform necessary action to ensure that no discrepancies occur.