Road Pothole Detection Based on Crowdsourced Data and Extended Mask R-CNN

Road pothole detection is significant for road maintenance which has been widely solved based on computer vision models in recent years. However, the accuracy of the pothole detection is still far from satisfactory and has the potential to be increased. One of the challenges is the diversity of the training samples. Another challenge is the extension of mature computer vision models. Therefore, in this paper, we created a pothole dataset with multiple road conditions and environments via samples of previous studies and developed an extended mask R-CNN model. Based on the created dataset, we compared the proposed model with the traditional model using the ResNet-50 and ResNet-101 networks. The experimental results show a mAP_0.5 of 92.1% on the test dataset, surpassing the traditional backbone networks ResNet-50 and ResNet-101 by 3.6% and 4.5%, respectively. Furthermore, we compared the details of the different models. The experimental results show that the total area prediction error is only 3.19% on the test dataset. It suggests that our model can precisely extract geometric features and the area information of detected potholes.