Felix Schill

Aerial SLAM with a single camera using visual expectation

Micro aerial vehicles (MAVs) are a rapidly growing area of research and development in robotics. For autonomous robot operations, localization has typically been calculated using GPS, external camera arrays, or onboard range or vision sensing. In cluttered indoor or outdoor environments, onboard sensing is the only viable option. In this paper we present an appearance-based approach to visual SLAM on a flying MAV using only low quality vision. Our approach consists of a visual place recognition algorithm that operates on 1000 pixel images, a lightweight visual odometry algorithm, and a visual expectation algorithm that improves the recall of place sequences and the precision with which they are recalled as the robot flies along a similar path. Using data gathered from outdoor datasets, we show that the system is able to perform visual recognition with low quality, intermittent visual sensory data. By combining the visual algorithms with the RatSLAM system, we also demonstrate how the algorithms enable successful SLAM.

Estimating ego-motion in panoramic image sequences with inertial measurements

This paper considers the problem of estimating the focus of expansion of optical flow fields from panoramic image sequences due to ego-motion of the camera. The focus of expansion provides a measurement of the direction of motion of the vehicle that is a key requirement for implementing obstacle avoidance algorithms. We propose a two stage approach to this problem. Firstly, external angular rotation measurements provided by an on-board inertial measurement unit are used to de-rotate the observed optic flow field. Then a robust statistical method is applied to provide an estimate of the focus of expansion as well as a selection of inlier data points associated with the hypothesis. This is followed by a least squares minimisation, utilising only the inlier data, that provides accurate estimates of residual angular rotation and focus of expansion of the flow. The least squares optimisation is solved using a geometric Newton algorithm. For the robust estimator we consider and compare RANSAC and a k-means algorithm. The approach in this paper does not require explicit features, and can be applied to patchy, noisy sparse optic flow fields. The approach is demonstrated in simulations and on video data obtained from an aerial robot equipped with panoramic cameras.

Effective Communication in Schools of Submersibles

Effective communication mechanisms in schools of submersibles are a key requirement for their meaningful deployment. Furthermore a fully distributed communication schema is preferable for reasons of reliability. The required communication form is usually many-to-many, especially omnicast [4] (or gossiping ). All these constraints are hard to achieve at the same time in a low-bandwidth, short range communication setup. The theoretical findings in [4] and [5] are expanded and employed for the actual underwater schools communication in the Serafina project (e.g. [2], [3]). The results are reported here.

Pruning Local Schedules for Efficient Swarm Communication

Reliable wireless communication underwater is a precondition for swarming technologies. This paper discusses a time division multiple access (TDMA) algorithm suitable for dynamic multi-hop wireless networks, which offers quick all-to-all information exchange (Omnicast), dense local schedules and predictable latencies. The algorithm is based on an earlier algorithm published by the authors in [Schill F., et al., 2006]. This paper presents an improved and simplified algorithm to calculate the local schedules, and uses a new mapping function for logical time slots to actual time slots, which balances sending frequencies between nodes. An extension of this algorithm is then presented which employs a technique to reduce the average degree of the connection graph as seen by the scheduling algorithm. It is explained how this reduction of degree can be achieved without causing communication collisions. The results of experiments performed in a real time simulation show the performance of the algorithm, and the performance gain achieved by local reduction of the degree.

Robust asynchronous temporal event mapping

Localisation and mapping relies on the representation and recognition of features or patterns detected in sensor data. An important aspect is the temporal relationship of observations in sensor data streams. This article proposes a new approach for simultaneous localisation and mapping based on temporal relations in the flow of characteristic events in the sensor data channels. A dynamical system is employed to acquire these correlations between simultaneous and sequential events from different sources, to map causal sequences, while considering time spans, and to recognise previously observed patterns (localisation). While this system is applicable to sensor modalities with different characteristics and timing behaviours, it is especially suitable for distributed computing. Mapping and localisation take place simultaneously in an life-long unsupervised distributed online learning process. The dynamical system was implemented as a distributed real-time system with symmetric processes. A real-time clustering network reduces the dimension of raw sensor data. Cluster transitions are used as input for the dynamical mapping system. Results from physical experiments with one sensor modality are presented.