AI-Based Rotation Aware Detection of Aircraft and Identification of Key Features for Collision Avoidance Systems (SAE Paper 2022-01-0036) 2022-01-0036
Object detection using deep learning is a well-studied area and different neural network architectures have been proposed for localization of objects at an eye-level view. However, detection of airplanes is more challenging as they are not necessarily aligned horizontally or vertically in the input images as is the case in vehicle or people detection. For aircraft detection, horizontal axis-aligned bounding boxes are not precise enough and may contain a plethora of background data. Thus, our approach for aircraft detection proposes to infer additional information about the orientation of the airplane directly from the object detection model. Additionally, we also apply a computer-vision post processing pipeline to find out the specific aircraft features such as tail, head, wings, etc. Combining the obtained angle and additional key features of the airplane allows for determining the direction of travel for aircraft which can be potentially used as a part or as an enhancement of more complex collision avoidance systems. Specifically, this study focuses on an in-depth evaluation of various deep learning-based solutions for the fully automated detection of the aircraft heading direction from multi-instance satellite imagery characterized by rich background and small spatial resolution of objects. The proposed approach was verified on open-source airplane datasets, proving its robustness, high accuracy, and its capability to generalize well to new image sets. Additionally, the presented technique has a potential for automated enhancement of existing datasets with additional information about object orientation or key points, eliminating the need for pixel-wise labelling which is beneficial for various future studies in the aerospace field.
Citation: Kwasniewska, A., Chougule, O., Kondur, S., Alavuru, S. et al., "AI-Based Rotation Aware Detection of Aircraft and Identification of Key Features for Collision Avoidance Systems (SAE Paper 2022-01-0036)," SAE Technical Paper 2022-01-0036, 2022, https://doi.org/10.4271/2022-01-0036. Download Citation
Author(s):
Alicja Kwasniewska, Onkar Chougule, Sneha Kondur, Sairam Alavuru, Rey Nicolas, David Gamba, Harsha Gupta, Dennis Chen, Anastacia MacAllister
Affiliated:
SiMa.ai, General Atomics Aeronautical Systems Inc.
Pages: 11
Event:
AeroTech
ISSN:
0148-7191
e-ISSN:
2688-3627
Related Topics:
Collision avoidance systems
Aircraft
Machine learning
Neural networks
SAE MOBILUS
Subscribers can view annotate, and download all of SAE's content.
Learn More »