Few-Shot Railway Intrusion Detection Without Forgetting via Double RPN and Detector
Tao Ye, Xiao Cong, Yuliang Li, Zhikang Zheng, Xiangpeng Deng, Xiangming Yan, Xiaosong Li, Xi Zhang · IEEE Transactions on Intelligent Vehicles · 2024
Railway intrusion detection is a serious task in intelligent rail transportation system, ensuring the safe operation of trains and the safety of people's lives. The existing railway intrusion detection methods mainly adopt the methods based on deep learning with large-scale datasets, which is undoubtedly costly. The emerging few-shot object detection method (FSOD) provides a potential solution to solve the suddenness and contingency of railway intrusion. However, there is an inherent contradiction between the performance of the novel class and base class in most FSOD, i.e., the model bias problem. To address the above-mentioned issues of railway intrusion detection, few-shot railway intrusion detection (FSRD), consisting of Balanced region proposal network (RPN), Double Detector and an efficient fine-tuning framework, is proposed to detect novel class by few-shot samples without forgetting base class. Balanced RPN is developed to generate high-quality proposals that are not biased towards any classes for subsequent training and inference. Double Detector is further proposed to handle the region of interest (RoI) features of the novel and base class respectively, and generalize the novel class without forgetting the base class. Upon these baselines, we also tailored an efficient fine-tuning framework to further improve overall performance. The proposed method is extensively experimented on our constructed few-shot railway intrusion dataset (FSRT2023). Experimental results show that the overall performance of FSRD significantly outperforms some state-of-the-art (SOTA) methods in all settings. FSRD can be used as a never-forgetting few-shot detector that learns rare novel classes in railway intrusion detection.