Efficient Scene Reconstruction for Unmanned Aerial Vehicles

There is demand for small, cost-effective Unmanned Aerial Vehicles (U A V s) in industrial applications. These machines are expected to perform autonomous decision-making tasks. Recent work shows promise for systems based on images or video. We are motivated to study Structure From Motion (SFM), a three-dimensional (3D) reconstruction technique from two-dimensional images. It uses low-cost cameras rather than expensive LIDAR such as laser scanning sensors. The goal of this work is to seek an efficient way to improve accuracy and performance of SFM on small UAVs that lack the computation needed to perform costly algorithms such as neural networks. A major step of SFM is to estimate the homography between two views of the same scene by matching pairs of distinct features in an image (also known as geometric verification). Accurate estimation of pairs leads to better 3D models. We study feature-based homography estimation with various Random Sample Consensus (RANSAC) algorithms. Specifically, pre-sampling is one of the most common strategies utilized to improve the original RANSAC algorithm. Inspired by the work of others, we implement a new pre-sampling RANSAC method. Our adaptation, Blur and Contrast RANSAC (BC-RANSAC), has state-of-the-art performance while reducing run-time by a factor of 6.78.