Vehicle Detection in Satellite Videos

Status: available
SupervisorRoman Pflugfelder

Problem Statement

The robust detection of specific objects such as buildings, ships and road networks has been an important challenge in remote sensing and satellite image processing. Since 2014, a new generation of satellites allow to extend the number of object categories to smaller objects such as vehicles, aircraft and many other objects. These satellites, e.g. WorldView, SkySat provide very high resolution panchromatic image data. This new instrumental opportunity innovates radically the field of surveillance with a multitude of potential applications from environmental surveillance, to border control, to traffic monitoring, just to name a few possible application fields.

One important challenge in satellite surveillance is the recognition of vehicles. The current resolution of commercial satellites is still too low to reliable detect cars from space. An increase in resolution is technically possible, but raises serious privacy issues. One way to overcome this dilemma is to use instead of single satellite images satellite video. The idea is to compensate the lack of spatial resolution with high frame-rate videos, hence with a higher sampling in the temporal domain. Satellite video has strong similarities to WAMI (Wide Area Motion Imagery). The task of this work is to apply standard approaches of object detection for WAMI to satellite video and to study the potential of vehicle detection from space.

Goal

Get familiar with satellite imagery. Compose WAMI datasets, train an existing deep neural network and validate the results with existing satellite videos. Implement the neural network for vehicle detection.

Workflow

The thesis can be combined with a preceding Informatik Praktika.

  • Review literature
  • Create training and validation dataset
  • Implement training and test algorithms
  • Test results on specific satellite videos
  • Optional: Improve algorithms for better results
  • Written report/thesis and final presentation

Requirements

  • Basic knowledge in computer vision
  • Basic experience in Matlab, C++, Python
  • Interest in Machine Learning, maths, statistics
  • Interest in GPU programming

Literature

  • R. LaLonde, D. Zhang, M. Shah. ClusterNet: Detecting Small Objects in Large Scenes by Exploiting Spatio-Temporal Information. arXiv, 2017.