Extracting curvilinear features from millimetre radar data

This paper describes an optimisation framework for fitting Bezier splines to noisy image features. The novel contributions are three-fold. Firstly, we describe how a template-based feature enhancement operator can be learned using a variant of the exponential correlation associative memory (ECAM). Secondly, we show how a uniformly initialised mesh of Bezier spline control points can be fitted to the resulting feature characteristics. Finally, we develop a mesh-pruning strategy that can be used to effect perceptual grouping of the resulting local splines into a quasi-global contour representation. The new framework for spline-fitting is demonstrated on the problem of extracting road structures from noisy millimetre radar data.