Faster R-CNN railway foreign body detection algorithm combined with attention between channels

The intrusion of foreign objects such as pedestrians and vehicles into the boundary of the railway has seriously threatened the safety of pedestrians and the safety of railway traffic. Aiming at the low utilization rate of the classic Faster R-CNN model for inter-channel information, in the feature extraction stage of the model, combined with the inter-channel attention module of SENet, the SE-Faster R-CNN model is proposed and used on this basis A more balanced loss function and adjustment of the size of the anchor improve the detection accuracy of the model. When the number of railway foreign body images is insufficient, first use VOC2007 to train the model, and then use part of the KITTI data set and a small amount of high-speed rail platform monitoring data and railway foreign body data to fine-tune the model, making the model better in the railway scene Performance. The experimental results show that although the algorithm speed has dropped slightly, the mAP of the improved Faster R-CNN model on the VOC data set has reached 80.3%, which is an increase of 6.9% compared to the original model. The image data on the railway platform and the railway foreign body The average accuracy on the image data reached 87%, especially for pedestrians and trains in the railway, the detection accuracy reached 89.61% and 90.9%, respectively. Studies have shown that the introduction of the SEnet structure into Faster R-CNN improves the feature extraction capabilities of the original model, which makes the model's detection accuracy for targets greatly improved, and can well complete the task of railway foreign body detection.