Problem Statement
Object detection is a very important and still unsolved problem in object recognition. For example, the problem becomes challenging in aerial imaging and remote sensing as the scenes and the data differ significantly from the case usually considered in computer vision.
The aim of this thesis is to study vehicle detection for a new sensor modality. Vehicles in satellite images are tiny (4-10 pixel), the recognition of vehicles is challenging as seen in the following image cropped from the original satellite image.
This thesis offers the opportunity to develop a vehicle detector based on short sequences of satellite images which are taken by Planet’s SkySat constellation. Motion can be an important cue to detect tiny objects.
Remote vehicle detection has potential to innovate traffic monitoring and traffic prediction models which are today essential for autonomous driving and smart cities.
Goal
Given our previous work on satellite video, research a new approach for vehicle detection with the All-Frames sequences of SkySat satellite images. Understand the use of spatiotemporal (recurrent), neural networks and associative memories and compare their differences for the given task. Study the potential impact of this new sensor modality. Participate with our international research team (IARAI, ESA Phi Lab, Planet).
Workflow
The thesis can be combined with an Informatik Praktika. Please, do not forget to attend the mandatory courses (as well for the Praktika). You will write a thesis proposal at the beginning where you are welcome to formulate your own research ideas in agreement with the supervisor.
Requirements
- Basic knowledge in computer vision or computer graphics (e.g. Master Programmes Visual Computing, Data Science)
- Basic experience in C++, Python, Julia
- Interest in machine learning, maths, statistics
- Interest in collaborating internationally with others