With specific flora and fauna, regional landscapes and forests constitute an important part of the cultural heritage. Several natural environments have already been classified as national or regional parks. The UNESCO World Heritage covers 13% of the protected forests in the world. Thus, preserving those sites represents a crucial issue. Such a safeguarding involves a detailed knowledge of the sites and forestry management plans. The management of a natural forest is traditionally based on forest plot inventories in which several features of the trees are measured. The set of data collected during these inventories represents the starting point of forest monitoring, flora preservation and risks prevention. Traditionally, measurements are made manually by operators. However, during the last decade, terrestrial laser scanning has become a new and promising way of measuring such attributes. This instrument provides a fine three dimensional point cloud virtual representation of the scanned scene. Trees location, stem diameter, and stem taper can be extracted from these point clouds using pattern recognition algorithms. In this paper we present a novel two steps way to improve the quality of tree branching detection in a three dimensional point cloud acquired by terrestrial laser scanner. This method was developed in order to enhance the results of a previous study. Our approach is based on the combination of a simplification step (using particle simulation), followed by a shape detection (discrete arcs of circle detection). It identifies the lack of accuracy in tree stem diameter measurements at branching junctions for further more detailed analysis.