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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Real-time object detection for railway freight vehicles faces challenges like scale variation and complex backgrounds.
• Existing Transformer-based detectors (e.g., RT-DETR) struggle with multi-scale representation and small object detection.

Purpose of the Study:

• To propose an enhanced real-time detection framework (DAF-DETR) addressing limitations in multi-scale and small object detection.
• To improve the efficiency and accuracy of detecting key components in railway freight vehicles.

Main Methods:

• Introduced a lightweight dynamic hybrid convolutional network with adaptive dynamic kernel allocation.
• Developed the PSM-DyT module integrating Pola linear attention, dynamic normalization, and frequency-domain enhancement.
• Implemented an improved Pyramid-IEL fusion module to address cross-scale feature imbalance.
• Utilized Inner-Shape-IOU loss for enhanced localization accuracy of small and irregular targets.

Main Results:

• DAF-DETR achieved substantial lightweight gains on a freight vehicle dataset.
• Demonstrated improvements in precision (1.2%), recall (2.7%), and mAP@50 (2.5%).
• Validated effectiveness and generalization on the VisDrone2019 dataset with a 2.2% mAP@50 increase.

Conclusions:

• The proposed DAF-DETR framework effectively enhances real-time object detection for railway freight vehicles.
• The method shows significant improvements in efficiency and accuracy, particularly for small and complex objects.
• The framework exhibits strong generalization capabilities across different datasets.