The road conditions detection using the convolutional neural network

Abstract

Poor road conditions present considerable obstacles for individuals, resulting in asset loss, bodily harm, and time inefficiency. Approximately 1.35 million fatalities are attributable to road traffic incidents. The Department of Public Works and Town & Country Planning conducted road surveys to assess and strategize maintenance efforts. The manual car survey requires additional time and an excessive budget. The automated system of artificial intelligence (AI) is widely recognized. This paper presents a model to detect road surface conditions utilizing video data. Four versions of convolutional neural networks (CNN) were utilized for this work. The model evaluation employed the mean average precision (mAP) measure. The video data was acquired via a smartphone mounted in a vehicle, comprising 10,984 photos for training and 2,198 images for testing. We trained and evaluated four versions of CNN architectures named YOLO, utilizing our data and GPU, with a specific emphasis on identifying cracks, potholes, and the condition of manhole covers. Of the architectures evaluated, YOLO V6 attained the greatest mAP score in comparison YOLO V5 to YOLO V8. The testing results with batch sizes of 4, 8, 16, and 32 are effective. The batch size of 32 yields the highest performance, achieving 87.38% mAP. Conduct the dropout normalization using rates of 0.25, 0.50, 0.75, and 1. The maximum mAP is observed with a dropout rate of 0.25, yielding a mAP of 85.40%. The model indicates that the government conducted road surface inspections with enhanced efficiency, enabling the planning of road repairs for public utility issues, which can lower transportation costs. Additionally, the model can be utilized to identify hazardous road conditions and minimize vehicular accident rates.