Machine Learning-based Indoor Localization for Micro Aerial Vehicles
Micro Aerial Vehicles (MAVs) are versatile platforms: their applications range from surveillance to search and rescue operations. However, MAVs have limited processing power due to their small size and cannot fall back on standard localization techniques in the indoor environment. To address this issue, an efficient on-board localization technique using machine learning was developed in the scope of this thesis. The vision-based approach estimates x, y-coordinates within a known and modifiable indoor environment. Its computational power is scalable to different platforms, trading off speed and accuracy. Histograms of textons—small characteristic image patches—are used as features in a k-Nearest Neighbors (k-NN) algorithm. Several possible x, y-coordinates that are outputted by this regression technique are forwarded to a particle filter to neatly aggregate the estimates and solve positional ambiguities. To predict the performance of the algorithm in different environments an evaluation technique is developed. It compares actual texton histogram similarities to ideal histogram similarities based on the distance between the underlying x, y-positions. The technique assigns a loss value to a given set of images, enabling comparisons between environments and the identification of critical positions within an environment. To compare maps before modifying an environment, a software tool was created that generates synthetic images to simulate those taken during an actual flight. We conducted flight tests to evaluate the performance of the approach. A comparison of the localization technique with the ground truth showed promising results: the algorithm has a localization accuracy of approximately 0.6m on a 5m×5m area at a runtime of 32 ms on board of an MAV. In a triggered landing experiment, the MAV correctly landed in or close to specified areas. The map evaluation technique was applied to various high-resolution images to identify suitable maps. The presented approach is based on three pillars: (i) a shift of processing power to a preflight phase to pre-compute computationally complex steps, (ii) lightweight and adaptable algorithms to ensure real-time performance and portability to different platforms, (iii) modifiable environments that can be tailored to the presented algorithm. These pillars build a foundation for efficient localization in various GPS-denied environments.
Faculteit der Sociale Wetenschappen