A rapid and precise detection and measurement method for road cracks based on video keyframes and DPPNet

Video detection must address issues of low accuracy, slow speed, and inability to accurately capture road crack information. To tackle these challenges, this article proposes a fast detection and measurement method for road cracks using video keyframes and DPPNet. First, a keyframe index space is designed based on multi-scale crack ridge features and change detection to enable rapid keyframe selection. In a second stage, by analyzing complex and dynamic detection scenarios, a high-precision semantic segmentation DPPNet is developed, incorporating dense partial convolution and pyramid modules. This design focuses on local aggregation of crack features across spatial and channel dimensions while reducing feature map redundancy. Finally, a method for measuring crack width in dynamic video under complex conditions is proposed. Results show that our approach outperforms existing methods according to standard indicators, Dice and MIOU have improved by 2.16 %–11.48 % and 1.17 %–5.58 % respectively, which can effectively improve the detection speed by 5–7 times, and meets millimeter-level width detection requirements.