As a novel method, the application of stereo vision on forest resources inventory will not only reduce the cost, but also brings even richer information. In this thesis, strategies on stereo measurement suitable for forest environment is discussed based on the summary of the application of stereo vision on other area. The focus of the thesis is on the abstract of tree features and the techniques of stereo matching.Firstly, the history of the development of computer vision and the Marr vision theory, which plays an important part in computer vision, is briefly introduced. Secondly, an introduction on the meanings of stereo vision on the s forest resources inventory is given after the summary on its successful application on other areas and the application perspective is looked forward. Moreover, the components of forest measurement system based on stereo vision is put forward.The image preprocessing on the tree photos taken in the field is introduced at first. In the FVision system, edges of tree stem can be detected by first order gradient operator, second order gradient operator or SUSAN operator respectively. The Canny operator is the best to the three edge detection criteria on detecting step edge. An LoG operator of large template (13×13)can detect the Zero Crossing tree edge of single pixel perfectly . Moreover, the moravec operator and the Forstner operator, which are very useful on detection of feature points in calibration and matching, are developed in FVision.The three coordinates systems in stereo vision and the pinhole camera model are introduced. The application of nonlinear error in measure model in the calibration of camera is further improved on the base of other researcher. By repeat the experiment over and over, a mistake in the formula is found in reference 76, which used to decompose the camera parameters by projective matrix. An one-lens calibration is developed successfully. With the camera parameters resolved by linear least square algorithm as the initials, the nonlinear error in measure model is used to optimize camera parameters. Photos with the size of 1280×960 are taken with the camera (FUJIFILM FinePix6900Zoom) moving between the two ends of the fixed bracket or with the camera bracket moved from one site to another. Then the camera parameters are resolved. The result shows that, during the moving the camera and photo taken at different sites, the resolved parameters of FUJIFILM FinePix6900Zoom camera meet the equation dx=dy, where dx or dy is stable between 0.0061 and 0.0062, while (u0,v0) is not stable. During the camera's slipping between the two ends of the fixed bracket, dx=dy , where dx or dy is also stable between 0.0061 and 0.0062 and (u0,v0) keep stable around (625,510). If only nonlinear error in measure model is used for two-lens calibration, the same conclusion as that of the one-lens calibration can be drawn for the camera's intrinsic parameters.Stereo matching is one of the most difficult problem in stereo vision. Based on the summary of current development of stereo matching, a stereo matching strategy fit for the tree edge is put forward after the attempts of several stereo matching algorithms which are successfully used in other areas. This strategy uses an interactive variable area stereo matching constrained by tree edge. Then, a global matching is done using nonlinear error in measure model. For the samestereo image pairs, the tree edge using as constrains are detected respectively by Canny operator with , LoG operator with a template size of 13×13 and Sobel operator. The measure of area similarity is taken respectively by entropy difference, invariant moment, correlation and the synthesis of the three. Different matching results are gained with different matching methods put forward in thesis.With the camera calibration result in chapter 3, the sign points are reconstructed. The camera slipping on the bracket, the baseline length increased from 50mm, when the baseline length reaches 400mm, the single reconstruction error stab