Cleared to land a multi-view vision-based deep learning approach for distance-to-touchdown prediction

Abstract

With the broader adoption of digital air traffic control towers, real-time video data is expected to complement the current surveillance system (if available) and improve airport performance in terms of safety and efficiency. However, to fully utilize such data, a suite of computer vision algorithms needs to be developed for extracting useful information from real-time video feeds. Currently, most of the studies in the literature have focused only on the detection and tracking of aircraft on the airport surface, while approaching aircraft also play an essential role in airport and runway operations. The distance-to-touchdown of approaching aircraft is a critical parameter in final approach spacing and departure sequencing. Therefore, this research proposes a deep learning approach for estimating the distance of approaching aircraft to touchdown using multi-view video feeds. The proposed approach adopts a state-of-the-art computer vision model with an auto-calibration technique for detecting the approaching aircraft and extracting feature vectors from multiple camera views under various lighting and weather conditions. Then, an ensemble approach is introduced for combining the input vectors for distance estimation. The approach is evaluated with both Changi Airport simulated and real video data. Firstly, the proposed approach is designed to be easily updated and adapted for different camera system configurations. Secondly, the proposed approach has successfully combined the strength of both monoscopic and stereoscopic approaches to provide accurate distance-to-touchdown prediction in various scenarios. The experimental results demonstrate the advantages of the proposed approach with stable performance and low predicted errors (Mean Absolute Percentage Error = 0.18%) in estimating the distance-to-touchdown up to 10 NM. Such capability in a Digital Tower environment can augment the runway controller’s sequencing and final approach spacing capabilities.