Fod Detection For Airport Runway Based On Faster R-CNN

Foreign object debris(FOD)at the airport,as a lethal killer that threatens the safety of the aircraft during the take-off and landing stages,can cause flight delays,takeoff interruptions,and even endanger the lives of passengers.At present,most of the foreign object detection systems at airport runways at home and abroad use radar-based detection and visual-assisted mechanisms.The detection results of small-sized FOD objects are not ideal.This paper proposes a vision-based FOD object detection system,using digital image processing methods and deep learning strategies,to accurately and efficiently detect small-sized FOD objects.Faster R-CNN is an important recognition algorithm in deep learning,which can achieve the high-level semantic features of images.The Faster R-CNN framework can effectively identify FOD objects that are difficult for general convolutional neural network algorithms to classify.This paper mainly researches the improved Faster R-CNN object detection algorithm,gives the FOD object recognition strategy,and compares and verifies the effectiveness of the algorithm and strategy on the FOD image database.(1)For large-sized FOD objects that often appear on airport runways,considering the effectiveness and practicability of the system,traditional digital image processing methods are used for change detection to identify more obvious FOD objects such as large screws and nuts;(2)For small-sized FOD objects that are not detected after the change detection process,recognition based on the Faster R-CNN framework is performed to generate object candidate regions and further classify and regression.Using Dense-Net instead of Res Net-50 for feature extraction greatly reduces network parameters.Improved classification loss function and regression positioning loss function in the RPN layer are proposed.Focal Loss is used to optimize the weights of the difficult-to-classify samples,so that the training results focus on small samples and small-sized FOD objects that are difficult to classify.Using KL Loss in the regression layer,combined with KL divergence,makes the network return and converge more effectively.Experimental results show that the method in this paper can achieve high real-time detection results,higher accuracy of FOD object detection,and good anti-interference.This is mainly reflected in the fact that Dense-Net can reduce network redundancy,mitigate the disappearance of gradients,and improve detection speed;using two new types of optimized loss functions can ensure the detection accuracy of classification and object positioning;The Faster R-CNN architecture improves the robustness of the detection module and ensures the system’s good anti-interference performance.Aiming at the five types of FOD objects(small steel balls,metal nuts,thin iron wires,large screws and small screws),the method in this paper achieves an accuracy of 98.96% FOD object detection,and the detection speed is more than doubled than the classic Faster R-CNN.