Pothole and patch detection on asphalt pavement using deep convolutional neural network

Resumo

The main obstacles to the widespread use of the PMS are the high financial and time costs for carrying
out on-site assessments and the difficulty of processing and analyzing the data to generate the diagnoses of the
current condition of the pavement. With technological advancement, some techniques such as computer vision,
image processing, and machine learning can automatically extract the information of the pavements' condition.
The present study proposes the exclusive use of images from cameras attached to a vehicle, simple collection and
reduced cost, and extraction of information on pavement defects using a CNN. The research developed object
detection models with YOLO architecture to identify potholes and patches. It was analyzed the metrics impact of
the image size (224x224, 320x320, 416x416 pixels) and number of iterations for Yolo version 3 and 4. As
expected, the increasing image size resulted in improved metrics results and the expansion of the iterations led to
an improvement in the IoU. The CNN that presented the best overall performance, combining all the metrics, was
based on Yolov3, with an image size of 416x416 and 6000 iterations training, in which it obtained an F1-score of
79.00%, an average IoU of 64.59%, and mAP@0.50 of 73.85%.