Research on key components for railway freight vehicles based on improved real-time object detection methods.

The detection of key components in railway freight vehicles is challenged by diverse object categories, large scale variations, blurred structural boundaries, and complex background interference. To address the limitations of existing real-time Transformer-based detectors such as RT-DETR in multi-scale representation and small object modeling, this study proposes an enhanced real-time detection framework. A lightweight dynamic hybrid convolutional network is introduced to replace part of the backbone, employing adaptive dynamic kernel allocation to preserve large receptive fields while significantly reducing computational cost and parameters. The proposed PSM-DyT module, an improved hybrid Transformer unit integrating Pola linear attention, dynamic normalization, and frequency-domain enhancement mechanisms, strengthens fine-grained feature extraction and edge texture representation for small-scale components in complex scenes. Additionally, an improved Pyramid-IEL fusion module, designed as an enhanced substitute for the baseline RepC3 structure, mitigates cross-scale feature imbalance and improves robustness under challenging environments. Moreover, the Inner-Shape-IOU loss, extending traditional IoU with internal structural consistency constraints, boosts localization accuracy for small and irregular targets. Experiments on a self-constructed freight vehicle dataset demonstrate that the proposed DAF-DETR achieves substantial lightweight gains while improving precision, recall, and mAP@50 by 1.2, 2.7, and 2.5%, respectively. Additional validation on the VisDrone2019 dataset yields a further 2.2% increase in mAP@50, confirming both the effectiveness and generalization capability of the proposed method.