Detecting Patterns in Vector Fields

We describe global and local methods for comparing flow fields, and a visualization tool that allows the user to adjust how flow fields are compared. Our first global method operates on path-lines and measures variations in orientation and curvature between samples on the same path-line. We then apply the global approach to first-order attributes computed from the gradient tensors of the vector field. We show that local distributions are useful when the goal is to identify distinct regions and visualize patterns of vector field attributes directly within the flow field. To aggregate local descriptors into a global signature that allows us to compare entire vector fields, we use clustering. Our experiments show that the global methods outperform the local methods based on vector spin images, while also being more efficient. In conclusion, we recommend the 1D global distributions that can be efficiently combined into a single measure of similarity with adjustable weights for each attribute. I. Introduction omparing vector fields will become increasingly crucial as computational methods for simulating fluid dynamics (CFD), sensor technology for dynamic data, and video surveillance increase in accuracy and ubiquity. Automotive applications of comparing vector field include engine design where engineers need to compare simulated in-cylinder flow to an ideal swirl flow 4 , and in traffic analysis where positioning and velocity data from cars may be used to identify emerging traffic jams after comparing with smooth traffic flow field. We present results from our research on: