Whereas in the last decades the acquisition and processing of archaeological ground-penetrating radar (GPR) data have become mature, the interpretation is still challenging. Manual delineation in three dimensions is time consuming, and often the determination of an isosurface value is not straightforward. This paper presents a method designed specifically for the extraction of buried linear features such as wall foundations, based on template matching. First, the three-dimensional (3D) GPR data cube is synthesized into a two-dimensional (2D) slice. To achieve this, an energy slice based on a sufficiently large time window may often be appropriate, although in this study a combination with other attributes, for example based on phase symmetry, made weak anomalies more distinct. In the next step, we compute the 2D normalized cross-correlation of the composite 2D slice and a number of templates with dimensions similar to the walls in the data set. Of the resulting correlation matrices, the highest correlation coefficient is kept for each pixel, if it exceeds a certain threshold. In this way, wall foundations are successfully mapped, but also many false detections are produced. The latter are greatly reduced in number by using a size threshold and discarding isolated features. The remaining regions are enclosed in bounding boxes, which after vertical extrusion can be used as a simplified 3D representation of the wall structures, and for the creation of a filtered isosurface. For the evaluation of our results, a manual interpretation was used. In the 2D case (i.e. when comparing the total area of the automatically mapped structures versus the manually delineated ones), both the detection rate and the correctness were~77%. Slightly lower rates (~71%) were obtained in the 3D case (i.e. comparing volumes). Our method was applied to the GPR survey of a Roman villa in Kent, UK.