Abstract: Abstract: In order to develop adaptive equipment for the mechanical harvest of fruit, a 3D model of a fruit tree was required for analyzing its dynamic properties which can help in tuning vibratory parameters. A Chinese hickory tree without leaves was selected for this study. The objective of this paper was to extract trunks of a fruit tree from images for 3D reconstruction based on a stereo vision with two images. All tree branches in an image were divided into two parts, up-side branches and down-side branches respectively, according to their different background. The up-side branches can be easily segmented by an auto-threshold binarization algorithm. The down-side branches were extracted by simultaneous tracking of the two edges. The branch in image can be treated as a series of scanning beams defined by two end points with a certain width. The objective of the tracking process was to find an optimal path from the seed scanning beam to the root scanning. At each tracking step, nm candidate scanning beams with certain costs were generated under edge constraints, and only n scanning beams with lowest cost were reserved for next tracking step. After successfully tracking, a branch can be extracted by filling its two edges obtained from the optimal path. Thinning the branch, a binary tree was employed to describe the topology structure of the tree. In view of the continuity of shape and width of the branch, intersection points and bifurcate points of the skeleton were repaired heuristically. Under the topology and epipolar constraints, branch correspondences in two images can be found easily. Likewise, for a pair of corresponding branches, points in one branch can be found by their corresponding ones in another branch by epipolar constraints. To improve the accuracy of a reconstructed 3D model of trunk, curve-based reconstruction with curvature constraint instead of point-based reconstruction was employed. Taking into consideration of the Pipe Model of botany, the radius of a real branch was regressed as a linear equation with the variable of its length. Finally, an example of reconstructed 3D model of tree created by Pro/E software was shown. The result demonstrated that the reconstructed 3D tree is realistic visually. However, some branches were not recovered due to heavy occlusion with others. To solve these problems, more images with different views of the tree are needed and a data-fusing algorithm should be provided to combine the obtained results. For a certain branch, a comparison of the estimated radiuses between the 3D reconstruction and the measurement was made, and the relative derivation was less than 9%. The main reason was the imprecise assumption that the real cross section of branch was not an absolute circle.