A new neural network for particle-tracking velocimetry

 We describe a new neural network designed to solve the correspondence problem of particle-tracking velocimetry. Given two successive pictures of marker-particles suspended in a fluid, it matches their images by approximately duplicating the fluid motion. We present the results of efficiency tests that reveal the excellence of its performance and its stability with respect to the presence of unmatchable particle images. We compare its success rate in image matching to that of the neural network of Grant and Pan (1995), and observe that it produces better results when the flows have more important changes in direction. It has the important advantages over the latter, of being better adapted to benefit from parallel computing, and of being self-starting, i.e. of not requiring to be taught about the fluid flow in advance.