Pavement distress detection with conventional self-driving car sensors

Abstract

Properly detecting asphalt pavement distresses before their expansion is a crucial task to characterize road condition in order to identify causes of deterioration and, therefore, maintain roads by selecting the appropriate intervention. Many automatic methods are developed to detect and assess pavement distresses. However, these methods face many limitations such as reliability, efficiency, and price. Hence, they are not frequently deployed for road inspections. In this project, we address those limitations by testing a new approach in collecting pavement data. We test self driving car’s 3D LiDAR and RGB camera in detecting pavement distresses. We build our data acquisition platform and installed it on a moving vehicle. We conduct tests on the 3D LiDAR to assess its performance. Results show significant discrepancy between the scanned point cloud and its reference. Moreover, the reconstructed surface from the scanned point cloud shows significant deformations. This is mainly due to the noisy measurements, the low accuracy especially on low reflective surface, and the low resolution of the scanned point cloud. On the other hand, RGB camera shows a better performance in detecting cracks. We introduce a full approach for crack detection. We use a non-expensive camera to collect data on a moving vehicle. We use a supervised learning algorithm to detect cracks. We adopt an original way in annotating the collected dataset. Finally, we process it with a region-based deep convolutional neural network for instance segmentation. We validate the trained model on 70 images, rich with different scenarios, crack shapes, and noises, collected from a moving vehicle with a lowcost camera. Results show the capacity of our approach in crack detection and segmentation. The precision of the trained model is 77.67%, its recall is 79.18%, and the average intersection over union is 64.6%.

Date	13 Oct 2020
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Maarouf Saad (Supervisor) & Gabriel J. Assaf (Co-supervisor)