Major airports worldwide have undertaken substantial expansion programs to accommodate the steady growth in air traffic, including new runways and taxiways. The complex airside operations use advanced surveillance systems, such as Advanced Surface Movement Guidance and Control Systems (A-SMGCS). This system provides four operational functions: surveillance, control, routing, and guidance. But the techniques used have drawbacks in terms of accuracy, cooperation, noise, and delay.
Moreover, suppose there are obstacles, such as buildings or metal objects, between the airplane and sensors. In that case, the plane’s measured position can differ from its actual location; they must control and manage aircraft speed, direction, and location on taxiways to minimize any potential risk of collision between different airplanes. A research team from NTU proposes a computer vision-based framework, Deep4Air, for airside surveillance for airport ground movements to monitor runway and taxiway areas for better airside safety management.
The primary motive is to improve safety by exploiting Convolutional Neural Networks (ConvNets) and high-resolution videos to detect aircraft. After detection, it is possible to track each aircraft in real-time. It can also estimate the aircraft’s location, speed, and distance. Finally, planes can be identified by combining camera and radar information. A network of up to 14 cameras is used to achieve the above. Among these 14 cameras, 3 to 7 cameras will cover a runway, and its vicinity is used.
Due to privacy issues, It isn’t easy to obtain good quality videos from digital towers. The Changi Airport videos were generated using the NARSIM simulator system to overcome privacy issues and created two separate groups with the same training and testing properties. Each group is containing seven different views of a video. All videos are FHD resolution (1920 ×1080) taken from an 80m-tall tower.
In this paper, the proposed framework for airside safety management with the team’s approaches for aircraft detection and tracking in runways and taxiways is as follows.
1. Overview Framework
Aside from the video input, it also requires the reference location of the airport taxiways, runways, and holding points. Speed and distance estimation could then be used to provide advanced warning assistance.
2. Input Data
To assign objects to their corresponding airport locations, they first defined those locations. This approach was taken as the input videos in our experiments are static; therefore, we only need to draw the locations once. Their position relative to the video frames does not change.
To transfer from a video pixel location to a geographic location, we need to calibrate each video. The simulator does not provide camera information, camera calibration. Instead, an ML model was trained to translate every pixel point to its corresponding geographic point using all-points’ geographic location.