UAV Camera Re-localization Based on Image Retrieval

Camera relocalization is a challenging problem in computer vision, particularly for unmanned aerial vehicles (UAVs) due to their mobility, limited resources, and harsh conditions. Although several methods exist for camera relocalization, they suffer from limitations such as indoor bias, limited accuracy and range, high computational cost, and single-camera assumption. This highlights the need for more efficient and robust camera relocalization methods for UAVs. To address this challenge, We propose a realtime UAV camera relocation system that utilizes an unsupervised deep learning framework to learn image relationships. Transfer learning from large-scale classified data enables faster training and higher accuracy. The system employs image retrieval and a convolutional neural network to match features from a single RGB image with location information in a database. It achieves strong rotational invariance, enhancing its performance and versatility in various scenarios. The system has demonstrated high accuracy and robustness in large outdoor scenes, with each image taking only 0.6 seconds to compute.