Airport Runway FOD Detection Based On Deep Learning

With the progress of the times,an increasing number of people are considering air travel,making aviation flight safety crucial.Due to the high speed of aircraft takeoff and landing,the presence of Foreign Object Debris(FOD)on airport runways can pose many potential dangers,leading to serious consequences.Currently,most domestic airport runways rely on traditional manual inspections for FOD detection,which is time-consuming and inefficient.Although there are some new detection technologies available,such as millimeter-wave radar detection,they are prohibitively expensive.With the rapid development of deep learning in recent years,computer vision-based mobile on-board runway FOD detection technology has made significant progress.In this thesis,we address the following issues related to mobile on-board runway FOD detection:(1)Due to the mobile nature of on-board FOD detection,there are variations in target scale and background,which have a certain impact on FOD detection accuracy.Additionally,FOD objects are often small and have limited texture features.To address the problem of target scale variation in mobile on-board FOD detection,this paper proposes a YOLOv7 network with multiple detection heads,including a detection head for small-scale targets to mitigate the impact of small-sized objects.Furthermore,the paper introduces Contextual Transformer Networks(CoTNet)into the backbone network and detection heads to incorporate global information and enhance the detection of large-scale targets,thereby reducing errors caused by target scale variation.To address the issue of FOD objects being mostly small with limited texture features,an attention mechanism is introduced to focus on the texture features of small FOD targets for feature extraction.(2)In small object detection algorithms,the deep network depth and a large number of parameters often result in poor real-time performance.To overcome the problem of poor realtime performance caused by a large number of parameters in FOD detection algorithms,this paper proposes an innovative lightweight design that applies lightweight processing to the improved network.The ELAN module in YOLOv7 is replaced with the C2 f module,which utilizes depth-wise separable convolution.This modification reduces the computational cost of the algorithm while maintaining accuracy,thereby improving real-time performance.(3)In practical applications,the PC-based algorithm has limited usability,and the direct deployment on PC incurs high costs and is limited to specific scenarios.To address the issue of poor usability in FOD detection algorithms,the PC-based algorithm is ported and deployed on mobile devices.A real-time FOD detection system is designed,which involves capturing video images on the image acquisition end and conducting FOD detection inference on the mobile end.Real-time alarms are provided through the display interface.Experimental results demonstrate that the proposed YOLOv7+CoTNet architecture,incorporating the attention mechanism,improves accuracy in mobile on-board FOD detection.The introduction of the C2 f module with depth-wise separable convolution effectively resolves the issue of poor real-time performance in the YOLOv7 framework.Lastly,the feasibility of the proposed approach is demonstrated through porting deployment and system design.