基于改进YOLOv5+DeepSort算法的块状磨屑的识别与计数研究

Research on the recognition and counting of block debris based on improved YOLOv5+DeepSort algorithm

传统块状磨屑的处理方式存在效率低、收集不及时等问题,为了实现块状磨屑的精准定位检测,需研究块状磨屑的识别及数量统计。文章在原始YOLOv5算法的基础上添加注意力机制后,进一步强化了块状磨屑的有效特征;用BiFPN结构代替YOLOv5原有PAN结构,加强了浅层特征的利用。添加大尺度检测层,可更加准确地定位小尺寸块状磨屑;选择EIoU作为目标框回归的损失函数,提高收敛速度。在此基础上,该研究采用DeepSort算法和虚拟检测线法实现块状磨屑的数量统计。利用轨道小车采集的数据进行块状磨屑的检测和计数试验。试验结果表明:较于原始YOLOv5算法,改进YOLOv5算法虽然检测速率有所下降,但是精确率提升了4%、召回率提升了7.5%、均值平均精度提升了9.7%,改善了小尺寸块状磨屑的检测效果。改进YOLOv5+DeepSort算法结合改进检测线在各场景下可较准确地实现对块状磨屑的计数。

The traditional treatment method for block debris has problems such as low efficiency and untimely collection.In order to achieve precise positioning and detection of block debris,it is necessary to study the identification and quantity statistics of block debris.This article further strengthens the effective features of block debris by adding CBAM attention mechanism on the basis of the original YOLOv5.Replacing YOLOv5's original PAN structure with BiFPN structure enhances the utilization of shallow features.Adding a large-scale detection layer can more accurately locate small-sized block debris.The paper selects EIoU as the loss function for the target box regression to improve convergence speed.On this basis,the DeepSort algorithm and virtual detection line method are used to achieve the quantity statistics of block debris.Detection and counting of block debris use data collected by rail cars.The experimental results show that compared to the original YOLOv5,although the detection rate of the improved YOLOv5 has decreased,the accuracy has increased by 4%,the recall has increased by 7.5%and the average accuracy has increased by 9.7%,improving the detection effect of small-sized block debris.The improved YOLOv5+DeepSort algorithm combined with improved detection lines can accurately count blocky debris in various scenarios.