An Efficient Few-Shot Object Detection Method for Railway Intrusion via Fine-Tune Approach and Contrastive Learning

Abstract

Railway intrusion detection plays an important role in the railway intelligent transportation system, assisting the safe operation of trains. The existing deep learning-based object detection methods excessively rely on a large number of supervision samples, and build up such a railway intrusion dataset is usually costly. The emerging few-shot object detection (FSOD) can realize object detection task by learning a small number of samples, which is a suitable candidate for railway intrusion detection. However, small databases lead to poor classification performances, and most existing FSOD methods have the drawbacks of high computational complexity and slow adaptation speed. To deal with these issues, an efficient few-shot object detection method for railway intrusion (FSRD) is proposed in this paper, extending FSOD to railway intrusion detection with few training samples to achieve good performance. A simple yet effective fine-tuning method is also developed to improve overall performance. There is a classifier head initializer to perform an appropriate initialization of the model weights and speed up the adaptation of the detector. The contrastive learning method is further introduced to cope with the misclassification. The proposed novel method is tested on a real railway intrusion dataset. Experimental results show that it achieves an average detection accuracy of 80.51 mAP50 under the setting of 30 shots. Compared with the state-of-the-art (SOTA) FSOD methods, it is capable of yielding the optimal and suboptimal performance in all settings. Furthermore, conducting extensive experiments demonstrates its good detection accuracy, fast adaptability, and satisfactory generalization ability in railway intrusion detection.