Robust FOD Detection using Frame Sequence-based DEtection TRansformer (DETR)

In this study, we develop a frame sequence-based transformer model for the automated detection of Foreign Object Debris (FOD) on airport runways. Our model integrates an LSTM network with a pre-trained DETR transformer to enhance detection robustness in terms of accuracy and consistence. Our approach captures short video sequences as input, using the encoder-decoder component of the DETR model to extract essential features. These features are then propagated through LSTM cells to incorporate temporal context. We explore various configurations of our proposed model and compare its performance with the baseline DETR. Experimental results demonstrate that our proposed model achieves significant improvements in both detection accuracy and consistency, showcasing its potential in enhancing safety on airport runways.