Personnel Intrusion Detection in Railway Perimeter with Improved YOLOv7

In the field of intelligent railway security, personnel intrusion within the railway perimeter is a dangerous behavior. However, detection of such personnel intrusion can be unsatisfactory due to various environmental factors and detection distance. Therefore, this article presents a comprehensive analysis of this problem and proposes three strategies to improve the accuracy and robustness of the detection model, which have achieved excellent results in recent times. Firstly, we replace the traditional convolution module in YOLOv7 with the SPD-Conv module, which applies no stride convolution or pooling layers to enhance feature extraction of small objects in the backbone network. Secondly, we introduce the SimAM attention mechanism in the neck network to increase the aggregation network’s attention to target features. Thirdly, we create a simple decoupling detection head for this specific single-category detection task. To verify the effectiveness of our proposed method, we created a railway perimeter personnel intrusion detection dataset from both real scenes and simulated events and conducted several comparative experiments. The results demonstrate our method significantly improves detection accuracy, and detects small objects with increased accuracy. Additionally, our method exhibits strong robustness in project field testing.