The paper discusses methodology and presents preliminary results of a low-cost method for visual tracking (in an exemplary setup consisting of a mobile agent, e.g. a drone, with an on-board that follows a mobile ground object randomly changing its location). In general, a well-known concept of keypoint-based image representation and matching has been applied. However, the proposed techniques have been simplified so that the future system can be prospectively installed on board of low-cost drones (or other similarly inexpensive agents). The main modifications include: (1) target localization based on statistical approximations, (2) simplified vocabulary building (and quantization of descriptors into visual words), and (3) a prospective use of dedicated hardware.
展开▼
