Content area
Culverts are hydraulic structures essential for managing the flow of water. The culvert asset management program (CAMP) aims to inventory the culverts administered and maintained by the New Mexico Department of Transportation (NMDOT) along state-maintained roads. The NMDOT estimates the presence of roughly 60,000 culverts. Field inspectors are bound to make mistakes when it comes to documenting each culvert for inventory. Hence, a quality control team was established to manually cross-reference each culvert’s photos with what a field inspector entered.
The current approach to reviewing culvert data presents a significant challenge to the manual quality control process. This thesis proposes the solution of artificial intelligence autonomously reviewing the inventory photos for culvert features and flagging discrepancies with the field inspector entries. The computer models are based on the most prevalent types of errors. The computer vision models will be trained on and applied to the most prevalent culvert types: concrete or corrugated metal culverts with circular or box openings, making up roughly 94% of the culverts inventoried by New Mexico Tech from 2021-2023. Three custom YOLO (You Only Look Once) computer vision models were developed based on convolution neural networks (CNN). The first model is an object detection model that analyzes the image from the shape of the culvert opening and counts the number of culverts in the image. The second model is an image classification model that recognizes the material of culverts in the image. Lastly, an image classification model is used to detect the silting conditions.
The image classification models for the material and silting models had a low loss value of 0.334 and 0.365 and a high accuracy of 0.980 and 0.984, respectively. The material model had a precision of 0.976, a recall of 0.988, and an F1 score of 0.982 when detecting concrete. The model also had a precision of 0.989, a recall of 0.971, and an F1 score of 0.978 for corrugated metal detection. The silting model had a precision of 0.971, a recall of 0.985, and an F1 score of 0.978 when detecting clean to moderate silting. The model also had a precision of 0.984, a recall of 0.969, and an F1 score of 0.976 for major silting detection. The object detection model had a low box loss of 0.958, and a low class loss of 0.594. The object detection model also achieved a high mAP50 of 0.9, a precision score of 0.922, a recall score of 0.816, and an F1 score of 0.870. All models scored high metrics indicating the models are healthy and are operating as expected. The computer vision models showed high accuracy, precision, recall, and F1 scores compared to previous applications of computer vision models in civil engineering provided in the literature review. Overall, the computer vision models show the potential to increase the total daily data reviewed by the quality control team.
The models are deployed on the Hugging Face Spaces platform, enabling the quality control team to increase the output of checked culverts per day, increasing efficiency and reducing errors in the database. This allows for the NMDOT to make better resource allocation and risk management decisions based on quality data.