Adaptive sampling in particle image velocimetry

Particle Image Velocimetry (PIV) provides a way to analyse the instantaneous fluid displacement within some field of view, without influencing the underlying fluid. The displacement is measured by comparing two images of the fluid separated by a small timestep, extracting the observed displacement between the imaged particles through the use of cross-correlation windows. The size and locations (grid spacing) of such are traditionally user-defined and have a significant impact on the accuracy of displacement measurement. To simplify the process, image analysis algorithms have, over time, reduced the number of controllable parameters through coupling to a small subset. This may accelerate the learning curve for users, yet greatly reduces the flexibility of the analysis process. Accordingly, optimising such parameters is user-dependent and often not possible due to unavoidable trade-offs between resolution and robustness. The work presented in this thesis, therefore, investigates methods to maximise the extraction of information from PIV images, autonomously and efficiently. To achieve this, the research was focussed in three areas; reduction of unnecessary computation, improving the flexibility of analysis, and adaptivity, i.e. automatic decision making. The thesis presents methods for each theme, analysing their performance and implications for optimal, automatic, PIV image analysis. The first approach is to optimise the existing analysis architecture, by preventing unnecessary computation based on local convergence of the displacement field, to achieve greater utilisation of computational resources. This is shown within to be an effective approach, reducing the number of correlations by almost half in typical conditions. Alternatively, the analysis algorithm can be made adaptive such that it can self-select better analysis conditions. This approach hinges on the de-coupling of analysis parameters, which is found to raise many questions, regarding the optimal choice of such parameters, that are clarified and explored within. One such question is the initial window size, for which a novel algorithm is presented which automatically selects the locally optimal value. Further to this, fully unstructured and semi-structured analysis algorithms are explored, with novel implementations presented in both cases. The semi-structured approach which is found to be simple, robust, and highly computationally efficient, while still providing much of the flexibility of fully unstructured, representing a promising avenue for future research.