A real-time automatic pothole detection system using convolution neural networks

Detecting a pothole can help prevent damage to your vehicle and potentially prevent an accident. Different techniques, including machine learning, deep learning models, sensor methods, stereo vision, the internet of things (IoT), and black-box cameras, have already been applied to address the problem. However, studies have shown that machine learning and deep learning techniques successfully detect potholes. However, because most of these successful attempts are peculiar to the location of the study, we found no study which has addressed the peculiarity of potholes in South Africa using a tailored-trained deep learning model. In this study, we propose using a convolutional neural network (CNN), a type of deep learning model, to address this growing problem on South African roads. To achieve this, a CNN model was designed from scratch and trained with image samples obtained from the context of the study. The classifier was adapted to distinguish between a binary class which identifies the presence or absence of potholes. Results showed a significant performance enhancement at a classification accuracy of 92.72%. The outcome of this study showed that this machine learning approach holds great potential for addressing the challenge of potholes and road bumps in the region and abroad.