城市轨道交通车载视觉YOLOv5轻量化目标检测研究

Research on Lightweight Object Detection of YOLOv5 Based on Vehicle Vision in Urban Rail Transit

YOLO系列算法因其独特的优越性而被广泛研究,但是复杂的模型使得其在某些应用场景受到了限制。针对城市轨道交通无人列车驾驶中对前方轨行区异物入侵检测的时效性,本文以YOLOv5s算法为基础,从主干网络轻量化设计和边界框回归损失函数两方面进行优化。首先,通过引入GhostNet网络来减少网络参数提升模型轻量化。然后,以DIOU作为边界框回归损失函数,提高了模型收敛速度和精度。以COCO数据集进行试验,结果表明,所提算法在模型轻量化和检测精度上相比YOLOv5s均有所提升。

YOLO series algorithm has been widely studied because of its advantages,but its complex model restricts itsapplication in some scenarios.In view of the timeliness of foreign object intrusion detection in the front track area during the driving of unmanned trains in urban rail transit,this paper,based on YOLOv5s algorithm,optimizes the lightweight design of the backbone network and the regression loss function of the boundary box.First,the GhostNet network is introduced to reduce the network parameters and improve the lightweight of the model.Then,DIOU is used as the boundary box regression loss function to improve the convergence speed and the accuracy of the model.The experimental results on COCO data set show that the proposed algorithm improves the model lightness and detection accuracy compared with YOLOv5s.