基于跨尺度特征融合的泵站安全帽检测方法

Safety Helmet Detection Method of Pump Station Based on Cross-Scale Feature Fusion

中文关键词:  泵站  安全帽检测  注意力机制  特征融合

英文关键词:pump station  safety helmet detection  attention mechanism  feature fusion

中文摘要:

      为应对泵站场景下设备和人员之间目标被遮挡及远距离小目标对泵站重点区域安全帽佩戴自动监管带来的挑战，提出了一种融合注意力机制和跨尺度特征融合的安全帽佩戴检测算法，以克服在远距离、有遮挡场景下安全帽检测准确度低的问题。通过采集泵站监控视频数据构建泵站场景安全帽数据集，在特征提取网络中加入注意力机制模块，使得模型更关注于小目标的通道信息；同时增加检测层使得特征融合时能结合多级特征，并使用柔和非极大值抑制（Soft Non-Manimum Suppression,Soft-NMS）和完全交并比（Complete Intersection over Union,CIoU)算法进行改进以减少遮挡目标漏检情况。在自建数据集进行试验，结果表明改进后的算法平均准确率达到93.5%，与其他目标检测算法相比精度均有所提升，证明该方法在泵站重点区域场景安全帽检测任务中具有良好的性能。

英文摘要:

      In order to deal with the challenges of target occlusion between equipment and personnel and the automatic supervision of helmet wearing caused by remote small targets in key areas of pumping station,a safety helmet detection algorithm combining attention mechanism and cross-scale feature fusion is put forward to overcome the problem of low accuracy of safety helmet detection in long-distance and occluded scenes.The pump station scene helmet data set is constructed by collecting the monitoring video data.And the attention mechanism module is added to the feature extraction network to make the model pay more attention to the channel information of small targets.At the same time，the detection layer is added so that multi-level features can be combined during feature fusion，Soft-NMS and CIoU algorithms is applied to reduce the missing detection of occluded targets.Experimental results on the self-built data set show that the average accuracy of the improved algorithm reaches 93.5%，which is higher than other target detection algorithms.It proves that the method has good performance in the safety helmet detection task in the key area of the pumping station.