Vanilla-YOLOv8 railway foreign object intrusion detection method based on feature redundancy reduction

DU Kaihua; XU Guiyang; BAI Tangbo

doi:10.11860/j.issn.1673-0291.20230157

Abstract：Online monitoring technology for railway foreign object intrusion plays a critical role in ensur-ing the safety of railway operations and the security of passengers'lives and property.To address the issue of incomplete and inaccurate detection of occluded targets and small targets in existing foreign ob-ject detection algorithms,this paper introduces a railway foreign object detection algorithm,Vanilla-YOLOv8,based on YOLOv8.First,leveraging VanillaNet's approach of reducing network depth,shortcut branches,and enhancing nonlinear capabilities through deep training strategy modifications and dynamic adjustment of activation function states,the proposed method mitigates problems like model degradation,time inefficiency,and the disappearance of low-level small target features caused by excessive network layers and shortcut branches.This enhances the model's feature extraction capa-bility and detection speed.Then,improved partial convolution is employed to reduce redundant fea-tures,ensuring optimal utilization of extracted features.Finally,a Squeeze-and-Excitation(SE)atten-tion mechanism is integrated into the network backbone to increase the weight of key features,enhanc-ing feature representation and detection capabilities for occluded and small targets.Experimental re-sults show that the Vanilla-YOLOv8 algorithm achieves the mean average precision of 98.7%,re-duces parameters by 61.39%,and reaches the recognition speed of 125 Frames Per Second(FPS).These improvements mark a substantial advancement over traditional image processing techniques in terms of speed and detection accuracy,offering a valuable reference for real-time online monitoring.