Resolution-adaptive Quadtrees for Semantic Segmentation Mapping in UAV Applications

Unmanned Aerial Vehicles (UAVs) have been employed for multiple tasks in the last decade, including planetary exploration. A common task for UAVs is mapping, with maps most commonly represented in variable-resolution form, such as Octrees or their 2D-counterparts, Quadtrees. Convolutional Neural Networks (CNNs) have revolutionized sensor data processing, but are heavily influenced by the training data and the resolution of the images. In addition, Octree and Quadtree-based map representations typically insert values into the deepest level of the tree, regardless of the position of the UAV and the resulting image resolution. To alleviate both these issues, we propose a novel Quadtree implementation that considers the resolution of the observations when calculating insertion indices. The probabilistic mapping approach considers the tree structure when looking up probabilistic priors to insert generated observations. The system is implemented in ROS and tested in simulation and experiments with vertical, horizontal and mixed motions at different frequencies of observations. The map reproduction is evaluated with the Kullback-Leibler divergence and qualitatively. Increasing the frequency and approaching a target for closer inspection is shown to improve the quality of the map reproduction. Results also indicate that low-frequency, noisy observations can be overcome by increasing the frequency, opening up avenues for further research on mapping CNN observations. For a video demonstration visit: https://youtu.be/FdKu8BzjK0I