Numerous methods have been proposed for the detection of single trees in airborne laser scanning (ALS) data. Most of them are highly dependent on the initial settings of the algorithm, as parameters (e.g. smoothing of the digital canopy height model) will affect the overall detection performance, and more particularly the trade-off between omission and commission errors. To tackle this issue, the use of prior information about the forest stand is possible when ground truth for the area is available. Alternatively, adaptive parametrization in the course of the detection procedure requires more complex algorithms which might have trouble when processing large areas. In this article a procedure for automated parametrization is presented. It is based on the unsupervised training of the detection algorithm with reference forest plots including coregistered field and ALS data. The local maxima filtering algorithm is adopted as it is simple and fast. The training step consists in evaluating the detection performance of the algorithm on the reference plots for several parameter combinations. Detection quality is evaluated as a trade-off between the number of correctly detected trees and the number of false detections. When trees are to be detected in a newly surveyed area, two possibilities for algorithm parametrization are compared. The first option is to use the parameter combination that is the more robust when used on the training set (“average” setting). The second option is to use the combination that yields the best detection on the ALS point cloud from the training set that best resembles the new data. The matching criterion is based on the Fourier spectrum of the canopy height model computed from the point cloud. 26 forest plots located in seven different ALS surveys of mountainous areas are used to test the workflow. Plots have a minimum area of 0.25 ha and represent various stand structures and tree species. To compare the two parametrization options and evaluate their sensitivity to the training set size (number of reference plots), a cross validation procedure based on repetitive sampling of the training set among the available 26 plots is performed. Results show that for training sets with less than 15 plots, the “average” setting performs better, whereas with training sets larger than 20, the matching procedure yields better detection performance. This method for unsupervised training is quite flexible as it can be used with any detection algorithm that requires initial parametrization. Moreover, the detection performance criterion can be modified in order to reflect the end-user preference regarding detection results. This study is an example of how single tree methods can benefit from an area-based analysis. Further work should investigate whether metrics usually computed for area-based methods (e.g. height quantiles) could also improve the point cloud matching.