Astro Drone: Using crowdsourcing to collect visual data for distance estimation

Keywords
Loading...
Thumbnail Image
Issue Date
2013-08-23
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Autonomous navigation of robots requires obstacle localization techniques. Obstacles can be localized based on visual data. Large datasets are needed to test methods for visual obstacle localization quantitatively. In this thesis, I describe how such a large dataset can be obtained using crowdsourcing. Contributors collected two types of data while playing a game with a camera-mounted model helicopter. First, Speed-Up Robust Features (SURFs) [1] were extracted from images taken while the helicopter approached an object. Second, the distance to the object (ground truth) was calculated on the basis of helicopter sensor data that is processed by a Kalman filter. I tested if the ground truth distance measurements to objects are valid and if the proposed method of collecting data using crowdsourcing is efficient. To test the validity of distance measurements, I showed that a custom measure based on SURFs correlated with the distance to an object on manually recorded data. When the measure was calculated on Astro Drone data, the same measure correlated with the ground truth distances. This indicates that the ground truth contains valid information about object distances. To test if crowdsourcing is an efficient way for collecting data, I compare the number of data samples that were collected by the game (718 in three months) to its development time (1 year). I conclude that crowdsourcing is not an efficient method for data collection in short-term research projects, but that long running projects can benefit from crowdsourcing because continued data collection eventually leads to large datasets.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen