Oriented Bounding Box Detection for Bird's Eye View Camera
Koskinen, Tomi (2024)
2024
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-03-15
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202403042669
https://urn.fi/URN:NBN:fi:tuni-202403042669
Tiivistelmä
Oriented bounding box detection has seen a lot of development and new methods in the recent years. Almost without exception, the detection target for these methods has been objects in aerial images, taken with drones or satellites. Generally object detection methods use axis-aligned horizontal bounding boxes to localize objects in an image, but object detection from aerial images benefits from using oriented bounding boxes, because they are able to surround the objects with less background clutter and less overlapping with other objects.
This thesis considers another similar perspective, that could also benefit from the use of oriented bounding boxes. This perspective, defined as bird’s eye view, can be achieved by using high up surveillance cameras pointed downwards towards the ground. Places like parking lots, container ports and other locations that can have densely packed objects like cars, containers, boats and trucks, can use object detection combined with software to better manage these areas and make operations faster and more efficient. A simple example would be to detect the number of vehicles in different areas of a huge parking lot, and then directing incoming traffic based on that information.
The main objective in this thesis were to find out answers for these research questions: Does the bird’s eye view object detection benefit from using oriented bounding boxes? Which open-source oriented bounding box detection method performs the best on bird’s eye view images? Are there some oriented bounding box detection method key attributes that are always better than the others? To answer these questions a small dataset of 820 images from the defined perspective was created. Six open-source implementations of oriented bounding box detection methods were trained and evaluated on this created dataset.
Rotated RetinaNet was evaluated to being slightly(1.9 percentage-points) more accurate while using oriented bounding boxes instead of horizontal bounding boxes. The fastest and most accurate method on the created dataset was YOLOv5+Circular Smooth Labeling(CSL), with a 0.405 mean average precision and 112.4 frames per second. This makes YOLOv5+CSL easy to recommend. The key attributes were: angle-free/based, anchor-free/based and single/two-stage. No correlation could be found between detectors’ key attributes and their performance on the created dataset.
This thesis considers another similar perspective, that could also benefit from the use of oriented bounding boxes. This perspective, defined as bird’s eye view, can be achieved by using high up surveillance cameras pointed downwards towards the ground. Places like parking lots, container ports and other locations that can have densely packed objects like cars, containers, boats and trucks, can use object detection combined with software to better manage these areas and make operations faster and more efficient. A simple example would be to detect the number of vehicles in different areas of a huge parking lot, and then directing incoming traffic based on that information.
The main objective in this thesis were to find out answers for these research questions: Does the bird’s eye view object detection benefit from using oriented bounding boxes? Which open-source oriented bounding box detection method performs the best on bird’s eye view images? Are there some oriented bounding box detection method key attributes that are always better than the others? To answer these questions a small dataset of 820 images from the defined perspective was created. Six open-source implementations of oriented bounding box detection methods were trained and evaluated on this created dataset.
Rotated RetinaNet was evaluated to being slightly(1.9 percentage-points) more accurate while using oriented bounding boxes instead of horizontal bounding boxes. The fastest and most accurate method on the created dataset was YOLOv5+Circular Smooth Labeling(CSL), with a 0.405 mean average precision and 112.4 frames per second. This makes YOLOv5+CSL easy to recommend. The key attributes were: angle-free/based, anchor-free/based and single/two-stage. No correlation could be found between detectors’ key attributes and their performance on the created dataset.