Zero-shot railway obstacle detection based on improved open vocabulary YOLO structure and multi-scale linear Mamba attention

The rapid advancement of intelligent transportation systems has heightened the importance of obstacle detection within railway boundaries. However, existing datasets for identifying irregular track targets remain scarce. These datasets are characterized by limited category diversity and insufficient data volume. Additionally, summarizing irregular targets in rail transit scenes poses significant challenges. Two primary issues hinder progress: First, models struggle to recognize objects when training categories undergo substantial changes. Second, enumerating obstacle categories is complicated due to their uncertain number. Moreover, current target detectors heavily rely on predefined training object categories. This reliance limits their applicability in open scenarios, particularly for open-domain irregular targets. Consequently, detection performance in such contexts remains unsatisfactory. To address these challenges, this paper proposes an open-vocabulary object detection method based on the multi-scale linear Mamba attention (MSLMA) mechanism and open-domain YOLO. Our approach integrates multi-scale dynamic convolution concepts to reconstruct linear attention in the model space. This optimization enhances the feature extraction process. Furthermore, we introduce an improved Wise-IoU loss function to boost detection accuracy for irregular targets. Experimental results validate the effectiveness of the proposed method. It demonstrates exceptional performance in detecting irregular objects in rail transit scenes, especially under complex backgrounds and multi-scale conditions. The method significantly outperforms existing open-domain approaches, contributing a novel framework for open-vocabulary object detection and providing a robust solution for obstacle detection in intelligent transportation systems.