基于改进YOLOv5s的地铁车辆螺栓定位检测

Bolt positioning detection for metro vehicles based on improved YOLOv5s

摘要:

螺栓连接是地铁中常用的连接方式,其长期受交变力的影响易产生松动等故障,若未及时检测则会危及行车安全。地铁车辆螺栓定位检测是螺栓松动识别的基础。传统的人工螺栓定位检测成本高且容易出现误检、漏检,以机器人视觉检测为代表的自动化检测逐渐成为主流。地铁车辆视觉检测采集的图片分辨率高、螺栓像素占比小且分布密集,通过改进YOLOv5s网络对螺栓进行定位检测。针对螺栓像素信息少且多次卷积造成特征信息损失的情况,提出一种特征加权融合模块(FWFM),用于融合骨干网络提取的各尺度特征图,并结合坐标注意力机制(CA),赋予不同尺度特征图相应权重,提高模型对螺栓的关注度；为了解决密集分布螺栓检测难的问题,用NWD-NMS替换YOLOv5s网络中的IOU-NMS, 更好地去除重叠的预测边界框。实验结果表明,改进后的 YOLOv5s 算法在螺栓定位检测中显示出良好的性能,AP@.5:.95达到 96%,相对于传统YOLOv5s 算法提高了46.4%。

Abstract:

Bolt connection is commonly used in the underground connection, its long-term impact of alternating forces are prone to loosening, if not detected in a timely manner will endanger the safety of travelling. Bolt positioning detection is the basis of bolt loosening identification. The traditional artificial bolt positioning detection is costly and prone to misdetection, leakage detection, robot vision detection as a representative of the automated detection has gradually become the mainstream. Metro vehicle visual inspection collects high resolution images, bolt pixels account for a small proportion and dense distribution, through the improvement of YOLOv5s network for bolt positioning detection. Aiming at the problem of small bolt pixel information and the loss of feature information caused by multiple convolution, a feature-weighted fusion module (FWFM) is proposed for fusing the feature maps extracted from the backbone network at various scales, and combined with the coordinate attention mechanism (CA), which assigns corresponding weights to the feature maps at different scales, to improve the model"s attention to the bolts. To solve the problem of dense distribution of bolts, NWD-NMS is used to replace IOU-NMS in YOLOv5 network, which better removes the overlapping prediction bounding box. The experimental results show that the improved YOLOv5s algorithm shows good performance in bolt localisation detection, AP@.5:.95达到 96%, which is 46.4% higher than the traditional YOLOv5s algorithm.