TOKYO—Toshiba Corporation (TOKYO: 6502) and Toshiba Digital Solutions Corporation have developed an AI for road surface anomaly detection that enables highly accurate real-time detection of potholes on the surface of expressways that may lead to a serious accident (Figure 1), and the effectiveness of this AI technology was verified by improving routine inspection of expressways by the Central Nippon Expressway Company Limited (NEXCO Central). Results were promising for real-world application of this pothole detection system using AI.
In a world first (*1), this AI for road surface anomaly detection, which was developed by Toshiba and Toshiba Digital Solutions, uses weakly supervised learning in the detection of potholes and predicts the position of an anomaly within the image after being trained using images labeled only with the presence or absence of potholes. Use of weakly supervised learning can reduce the time required for preparing the training data to approximately 1/100th of that needed for a conventional approach (approximately 1 sec per image for weakly supervised learning vs approximately 1 min 40 sec per image for a conventional approach, Figure 4). This reduces the workload for the introduction of this AI and facilitates its implementation for the inspection of various roads.
A joint verification experiment conducted with NEXCO Central demonstrated the effectiveness of this AI technology for achieving highly accurate real-time detection of potholes on images acquired by a camera installed on a NEXCO Central vehicle while driving. This AI facilitates automation and labor-saving for routine inspection of expressways, and achieves early detection of potholes requiring urgent repair, thereby contributing to maintaining the stable operation of expressways.
Toshiba will present the details of this AI and the verification experiment on September 12 at PHMAP23, an international conference on infrastructure conservation to be held in Tokyo from September 11.
- HP ZBook Studio 15.6-inch G8 Mobile Workstation/Intel® Core™ i7-11800H Processor (maximum frequency, 4.60 GHz; total cores, 8/total threads, 16; cache, 24 MB), 32 GB DDR4-3200 memory (on board), NVIDIA® GeForce RTX™ 3070 (8 GB GDDR6)
- Measured by ROCAUC, an index used for evaluating the anomaly detection algorithm. The false positive rate (x-axis) was plotted against the true positive rate (y-axis) at various threshold settings, and the percentage of the area under the curve was calculated. A higher percentage indicates higher accuracy.