Enhancement Methods to Railway Intrusion Detection Problem Based on Video Surveillance Data
Jinghua Zhou, Bin Yue, Hu Chen, Shuo Ouyang, Jinchuan Chai · 2024
In recent years, video surveillance has been increasingly used in railway intrusion detection, but challenges still exist. First, the accuracy of intrusion detection is limited due to the limited samples of intrusion video frames. Second, different railway scenes are captured under different conditions, which increases the complexity of detection. In addition, similar backgrounds make it difficult to distinguish video frames captured by a single camera. To address these problems, we propose an effective few-shot learning algorithm. This algorithm utilizes processed railway video frames for training through three enhancement methods (base model improvement, attention mechanism, and image enhancement) in order to build a smarter model. The results show that the improved algorithm can successfully adapt to new scenes with only a small number of new samples and has better accuracy in intrusion detection.