Improved RT-DETR framework for railway obstacle detection

Abstract

Obstacle intrusion detection in railway systems is a critical technology for ensuring the operational safety of trains. However, existing algorithms face challenges related to insufficient multiscale object detection, high model redundancy, and poor real-time performance. Building upon the RT-DETR framework, this study proposes a Multiscale Separable Deformable (MSD) module that integrates depthwise convolution with deformable convolution to enhance feature extraction capabilities while reducing computational load. Additionally, a Deformable Agent Attention (DAA) mechanism is designed to optimize attention weights through sparse queries, effectively improving detection accuracy for small targets and enhancing inference speed in complex scenarios. Experimental results demonstrate that the improved model achieves 87.9% mean average precision (mAP) on a railway dataset, with a detection speed of 90 frames per second (FPS). The proposed model achieves a +1.7% mAP improvement and 13.9% faster inference speed compared to RT-DETR, while simultaneously reducing model parameters by 24.6%. As a result, the proposed model is highly effective for multiple obstacle intrusion detection in complex real-world scenarios.