基于改进YOLOv4的地铁车辆螺栓状态检测方法

Improved YOLOv4 detection algorithm for metro vehicles bolt condition

针对地铁车辆螺栓数量多、目标小导致的高误检率问题,提出一种基于改进YOLOv4的地铁车辆螺栓状态检测方法。采用级联的策略分2步检测实现螺栓目标的定位与分类,从而降低误检率。两步检测算法均采用改进的YOLOv4:首先利用聚类算法获取先验框的尺寸,并通过改变先验框生成的初始位置,优化预测框位置回归的策略,以提升网络训练与检测速度,其次重构了特征融合机制,用自适应空间特征融合的方法对PANet模块的输出进行操作,改善了特征的比例不变性,提高了对螺栓的检测精度。实验结果表明,提出的检测方法mAP、召回率分别可达99.5%,99.8%,可更高效地分类与识别螺栓小目标。

Aiming at the problem of high error detection rate caused by large number of bolts and small target in metro vehicles, a detection method for the status of bolts in metro vehicles based on improved YOLOv4 is proposed.In order to reduce the error detection rate, a cascade detection strategy is adopted to realize the positioning and classification of bolt targets in two steps.Improved YOLOv4 is adopted for both detection algorithms: First, clustering algorithms are used to obtain a priori box size, and by changing the prior box to generate the initial position. After that, the box position regression forecasting strategy is optimized to enhance the speed of network training and testing, and the feature fusion mechanism is reconstructed. Based on the aboved, by the method of adaptive spatial feature fusion PANet module output, the characteristics of scale invariance is improved, the detection accuracy of bolts is improved.The experimental results show that the proposed detection method mAP and recall rate can reach 99.5% and 99.8% respectively, which can classify and identify bolt small targets more efficiently.