Abstract The paper describes the onboard vision-based navigation software employed during the DLR's AVANTI experiment and presents its flight performance. The navigation task relies on a dedicated target detection module in charge of extracting the line-of-sight observations from the images taken by a monocular camera. The recognition of the target is based on the kinematic identification of its trajectory over a sequence of images. Subsequently, the resulting angles-only measurements are validated dynamically and processed by an Extended Kalman Filter to derive in real-time the relative state estimate. A computationally light implementation of the filter is achieved by the use of an analytical relative motion model. Two autonomous rendezvous to a noncooperative object have been performed in orbit, first from 13 km to 1 km separation, then from 3 km to 50 m. Despite the poor visibility conditions and the strong orbit perturbations encountered at low altitude, the filter was able to supply the onboard guidance and control algorithms with accurate and reliable relative state estimation, enabling thus a safe and smooth approach.