Improved YOLOv8 track foreign object detection based on lightweight convolution and information enhancement

Railroads are among the most common modes of transportation globally, but their tracks are typically open, making them vulnerable to the sudden and random appearance of foreign objects. Consequently, railroad accidents resulting from these obstacles have become increasingly common in recent years. This paper aims to efficiently and rapidly detect foreign objects on tracks to safeguard lives and property. Firstly, to reduce the model size, we introduce a lightweight convolutional module, AMConv. Additionally, to enhance the distinction between background and foreign objects, we propose a Contextual Information Enhancement Module (ClEM) in the backbone network to extract more pertinent feature information about foreign objects from the original image. Lastly, to enhance the accuracy of detecting foreign objects with small profiles, we remove the largest detection head and instead incorporate a smaller-sized detection head. Through numerous experiments conducted on the dataset, alongside ablation and comparison experiments with state-of-the-art models, the effectiveness of our proposed model is demonstrated.