講演情報
[Data_Collection-18]Development of a Crack Evaluation Method for Expressways Using Deep Learning and Comparative Verification with Conventional Methods
*Keizo Endo1, Shigeru Shimeno1, Tetsuya Konno1, Shinpei Chiba1, Hayato Nakajo1 (1. Nexco-East Engineering Company Limited (Japan))
キーワード:
Deep Learning、Expressways、Crack Detection、Neural Network
Research Objective
Crack evaluation in pavement surface condition surveys has traditionally been performed visually by skilled inspectors, which takes significant time and often delays reporting. Therefore, we developed an automated crack detection system using deep learning to ensure evaluation accuracy and improve operational efficiency.
Method
High-performance pavements with large service areas on expressways—Type I (porous asphalt pavement) and Type II (similar to stone mastic asphalt pavement)—are relatively difficult for crack detection. To address this, we developed a crack detection algorithm using the following approach:
Conventional Method: Road surface images captured by a line-sensor camera mounted on a pavement condition measurement vehicle were visually inspected by inspectors to determine the number of cracks and the state of patching. This process required approximately 45 minutes per kilometer per lane.
New Method: A neural network was applied to classify images into four crack categories for each block.
・The learning models “Deep Lab v3+” and “ResNet50” were adopted, and approximately 13,000 labeled images were used as training data.
・For crack evaluation, objects outside the scope—such as bridge joints and lane markings—were automatically extracted and excluded from the evaluation blocks. Patching was also automatically extracted beforehand and included when calculating the crack ratio.
・Cracks were then traced, and the crack ratio was calculated for each block based on the traced cracks.
Results
The automated analysis system reduced the time required for crack evaluation and crack ratio calculation to approximately 10 minutes per kilometer per lane. Furthermore, compared with visual evaluation based on expressway surface images, the system achieved a high detection accuracy of about 95%.
Academic Contribution
We developed an AI-based crack detection algorithm for expressway pavements that ensures accuracy comparable to conventional visual evaluation while achieving efficient automation of crack assessment.
Crack evaluation in pavement surface condition surveys has traditionally been performed visually by skilled inspectors, which takes significant time and often delays reporting. Therefore, we developed an automated crack detection system using deep learning to ensure evaluation accuracy and improve operational efficiency.
Method
High-performance pavements with large service areas on expressways—Type I (porous asphalt pavement) and Type II (similar to stone mastic asphalt pavement)—are relatively difficult for crack detection. To address this, we developed a crack detection algorithm using the following approach:
Conventional Method: Road surface images captured by a line-sensor camera mounted on a pavement condition measurement vehicle were visually inspected by inspectors to determine the number of cracks and the state of patching. This process required approximately 45 minutes per kilometer per lane.
New Method: A neural network was applied to classify images into four crack categories for each block.
・The learning models “Deep Lab v3+” and “ResNet50” were adopted, and approximately 13,000 labeled images were used as training data.
・For crack evaluation, objects outside the scope—such as bridge joints and lane markings—were automatically extracted and excluded from the evaluation blocks. Patching was also automatically extracted beforehand and included when calculating the crack ratio.
・Cracks were then traced, and the crack ratio was calculated for each block based on the traced cracks.
Results
The automated analysis system reduced the time required for crack evaluation and crack ratio calculation to approximately 10 minutes per kilometer per lane. Furthermore, compared with visual evaluation based on expressway surface images, the system achieved a high detection accuracy of about 95%.
Academic Contribution
We developed an AI-based crack detection algorithm for expressway pavements that ensures accuracy comparable to conventional visual evaluation while achieving efficient automation of crack assessment.
