The various elements of a modern target tracking framework are covered. Background theory on pre-processing, modelling and estimation is presented as well as some novel ideas on the topic by the author. In addition, a few applications are posed as target tracking problems for which solutions are gradually constructed as relevant theory is covered. Among considered problems are how to constrain targets to a region, use state-independent measurements to improve estimation in jump Markov models and how to incorporate observations sampled at an uncertain time into a state-space model. A framework is developed for tracking dolphins constrained to a basin using an overhead camera that suffers from occlusions. In this scenario, conventional motion models would suffer from infeasible predictions outside the basin. A motion model is developed for the dolphins where collisions with nearby walls are avoided by turning. The basin is modelled as a polygon where each point along the edge influences the turn rate of the dolphin. The proposed model results in predictions inside the basin, increasing robustness against occlusions. An extension to a Gaussian mixture background model providing a degree of confidence for detections is used to improve tracking in the presence of shadows. A probabilistic data association filter is also modified to estimate the dolphin extension as an ellipse. The proposed framework is able to maintain tracks through occlusions and poor light conditions. A framework is developed for estimating takeoff times and directions of birds in circular cages using an overhead camera. A jump Markov model is used to model the stationary and flight behaviours of the birds. A proposed extension also incorporates state-independent measurements, such as blurriness, to improve mode estimation. Takeoff times and directions are estimated from mode transitions and results are compared to manually annotated data. The cameras are inaccessible in both applications, disallowing proper calibrations. As an alternative, a method is proposed to estimate stationary camera models from available data and known features in the scene. A map of the basin and the funnel dimensions are used respectively. The method estimates a homography and distortion parameters in an invertible mapping function. An extension to the linear Gaussian state-space models is proposed, incorporating an additional observation with an uncertain timestamp. The posterior distribution of the states is derived for the model, which is shown to be a mixture of Gaussians, as well as some estimators for the distribution. The effects of incorporating the observation with an uncertain timestamp into the model are analysed for a one-dimensional scenario. The model is also applied to improve the GPS position of an orienteering sprinter by using the control position as an observation with an uncertain timestamp. / LINK-SIC
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-132426 |
| Date | January 2016 |
| Creators | Veibäck, Clas |
| Publisher | Linköpings universitet, Reglerteknik, Linköpings universitet, Tekniska fakulteten, Linköping |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Licentiate thesis, monograph, info:eu-repo/semantics/masterThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
| Relation | Linköping Studies in Science and Technology. Thesis, 0280-7971 ; 1761 |
Page generated in 0.0021 seconds