Potholes Detection from Dashcam using Deep Learning Approach

Deep learning has attracted considerable interest in the previous ten years and has established itself as a leading technology in the artificial intelligence sector. In object detection based on image processing, characteristics are extracted from images, and after that, data including category, position, and motion are collected and analyzed. Object detection same goes with potholes detection. Due to the potential for crashes and injuries, potholes can be dangerous for cyclists, pedestrians, and moving automobiles. The discovery and repair of potholes may fundamentally change with the development of automated pothole detecting systems. Road repair is necessary to prevent such traffic accidents in the future. Therefore, in order to perform road maintenance, it is necessary to identify difficulties with things like potholes. The aims of this research is to identify number of potholes on asphalt or concrete roads from the images obtained for both primary and secondary data, to investigate the accuracy of the testing by using the deep learning approach which is Convolutional Neural Network (CNN) algorithm and to differentiate between normal road and potholes by evaluating them using confusion matrix which are “actual potholes”, “predicted potholes”, “actual normal” and “predicted normal”. Data collecting has been done around Universiti Teknologi Malaysia's road. Deep learning has been applied in this study is CNN (Convolutional Neural Networks), which was used as the model of this project. Results obtained are in confusion matrix which it concludes the actual number of potholes there. The actual potholes obtained were less than expected so this means more data needed if this project extended in the future. To improve and to achieve better performance of this project, data collecting can be done by adding other places so there will be more data and resolution of the images can be sharper so the results may be more accurate. To make our roadways reliable, efficient, and long-lasting, research on pothole detection is essential.