A Detection Confidence-Regulated Path Planning (DCRPP) Algorithm for Improved Small Object Counting in Aerial Images
Miles Krusniak, Leppanen Keaton, Zhicheng Tang, Fan Gao, Yizhen Wang, Yi Shang · 2020
Computerized object counting shows potential for conservation population estimates as an alternative to manual counting, but remains unsatisfactory for minuscule objects, such as those in UAV-produced images. We improve UAV data collection by using a novel path planner which shifts altitude to maximize deep learning-based object detection confidences from a Faster Region-Convolutional Neural Network, considering energy consumption trade-offs. Using an empirical altitude confidence relationship, our adaptive path planner (”DCRPP”) adjusts UAV altitude based on confidence, yielding better quality data given energy constraints. DCRPP achieves 11.92% greater accuracy compared to fixed-height methods in our conservation-aimed simulation.