Track fastener detection in special scenarios based on TSR-Net

Abstract

The traditional method of detecting track fasteners can lead to erroneous results due to the presence of rail bolts, wires, and extraneous objects such as stones. Consequently, the study of fastener detection in specific environments is essential. To address these issues, the TSR-Net target detection model is introduced, which employs an efficient vision transformer as a backbone to increase the speed of feature extraction. The global image perception of the network is enhanced by integrating translation convolutions and inverted residual blocks within the neck network. A detection head with a self-supervised equivariant attention mechanism is designed to deal with the occlusion challenges of small fastener targets. The model is implemented on edge AI computing devices. Technical validation indicates that TSR-Net achieves 94.2% detection precision and operates at 47 frames per second, thereby enabling accurate, real-time detection of small-target occlusion fasteners.