Sim2world: few-shot domain adaptation for object detection
Advancements in Digital Tower Technology offer significant benefits to air traffic control and airport operations and management. Object detection models play a key role in identifying various aircraft, objects, vehicles, and personnel in the airport environment using visual data such as surveillanc...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2024
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/177717 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Advancements in Digital Tower Technology offer significant benefits to air traffic control and airport operations and management. Object detection models play a key role in identifying various aircraft, objects, vehicles, and personnel in the airport environment using visual data such as surveillance cameras. These models, based on deep learning techniques, enhance air traffic controllers' awareness, thereby assisting them in achieving safer air traffic management. However, data scarcity in airport environments due to security concerns often leads to poor detection performance when using models trained on related but different domains. To address this issue, this project proposes and explores a method that combines state-of-the-art domain adaptation techniques, Test-Time Augmentation (TTA), and temporal information adjustment. The method aims to create an accurate domain-adapted object detector using a few-shot samples from the target domain, capable of localizing aircraft from images or video streams and classifying them based on their family type (e.g., A320, B737). Images from a flight simulator and Narita Airport surveillance cameras were collected, representing source and target domains, respectively. The performance of the adapted object detector was found to improve significantly, with an increase from an unadapted object detector mean average precision (mAP) of 0.1158 to 0.7645 mAP when 10 shots per class were used for adaptation. However, TTA and temporal information adjustment did not have a significant impact on the performance of the object detector. |
---|