FOD Detection using DenseNet with Focal Loss of Object Samples for Airport Runway

Abstract

The foreign object debris (FOD) on the airport runways is a serious threat to aircraft taxiing, taking off and landing. The installation of cameras on the side of the track to automatically monitor the FOD is very important for improving the safety of the runway. This paper proposes a FOD detection method based on Convolutional Neural Network (CNN) for small-scale FOD. This method employs FasterR-CNN framework to generate candidate regions for the input image. This paper uses DenseNet instead of the traditional VGGl6Net for feature extraction, which can greatly reduce the network parameters and is beneficial to small-scale FOD detection. This paper also modifies the loss function of classification in RPN and uses Focal Loss to optimize the weights of positive and negative samples, so that the training results are focused on FOD samples with high difficulty in classification in dense samples.Experiments show that our proposed method has good performance in real-time, detection accuracy and anti-interference. The airport runway FOD image dataset mainly includes four types of objects (small steel balls, metal nuts, large screws, and small screws). Compared with the classical FasterR-CNN, our method achieves 95.60% FOD detection accuracy with an increase of 15.91%, while the detection speed is also more than doubled.