Railway Traffic Object Detection Using Differential Feature Fusion Convolution Neural Network
Tao Ye,Xi Zhang,Yi Zhang,Jie Liu
DOI: https://doi.org/10.1109/tits.2020.2969993
IF: 8.5
2021-03-01
IEEE Transactions on Intelligent Transportation Systems
Abstract:Railway shunting accidents, in which trains collide with obstacles, often occur because of human error or fatigue. It is therefore necessary to detect traffic objects in front of the trains and inform the driver to take timely action. To detect these objects in railways, we proposed an object-detection method using a differential feature fusion convolutional neural network (DFF-Net). DFF-Net includes two modules: the prior object-detection module and the object-detection module. The prior module produces initial anchor boxes for the subsequent detection module. Taking the initial anchor boxes as input, the object-detection module applies a differential feature fusion sub-module to enrich the sematic information for object detection, enhancing the detection performance, particularly for small objects. In experiments conducted on a railway traffic dataset, compared with the current state-of-the-art detectors, the proposed method exhibited significant higher performance and was more effective and more efficient than the other methods for object detection in railway tracks. Additionally, evaluation results based on PASCAL VOC2007 and VOC2012 indicated that the proposed method was significantly better than the state-of-the-art methods.
engineering, electrical & electronic,transportation science & technology, civil