KNOWLEDGE GENERATION AND TACTICAL SMALL OBJECT DETECTION

Abstract

This research paper delves into the realm of railway safety, presenting a novel application of artificial intelligence to enhance real-time detection and prevention of potential rail track incidents. The study primarily focuses on two critical areas of object detection - rail track detection and tactical small object detection. Our customized dataset is based on real-world 4K video footage, capturing smaller objects like humans and miscellaneous debris present on the rail tracks that could lead to catastrophic accidents or derailments. In this research, we propose an innovative approach by employing the YOLOv5 model for accurate rail track detection and the application of a Global-Local Self-Adaptive Network (GLSAN) for efficient tactical small object detection. GLSAN significantly leverages attention mechanisms and multi-scale feature fusion, thus providing superior detection performance for small objects. Further, this study introduces the concept of 'knowledge generation' in object detection, using the metadata generated during the detection process to anticipate potential safety threats and take proactive safety measures. The outcomes of this study emphasize the efficacy of the proposed method, reflecting impressive accuracy and precision-recall values. This work promises a substantial contribution to the railway industry's quest for incident-free operations by potentially mitigating risks and enhancing railway track safety. Future directions for this research include refining the system's real-time performance and integrating multi-modal sensor data to further improve system robustness