Abstract: Aeolian saltation plays a very key role in shaping varieties of landforms in arid area and in affecting global climate system and marine ecosystem. It is generalized as four sub-processes: aerodynamic entrainment, the grain trajectory, the grain-bed collision and wind modification. Among these sub-processes, individual particle trajectory formation is a key chain affecting grain-bed collision and interaction between sand particles and wind conditions. Nevertheless, up to date, the mechanism of sand particle trajectory formation has not yet disclosed perfectly due to lack of a sophisticated algorithm for extracting effectively sand particles from high-concentration images. The traditional algorithms with single thresholding often cannot segment sand particles effectively from backgrounds due to brightness difference among saltating particles, stable and stochastic noises in high-concentration images. The algorithm with dynamic thresholding proposed by Ohmi and Li, however, needs to preset arbitrarily empirical parameters as maximum, minimum and contrast threshold, probably introducing uncertainty of segmentation. Thus, an improved segmentation- algorithm with dynamic thresholding is proposed here, which covers denosing by substracting a background image, graying by green channel, differentiation, targets’ detection by gray-level variance and segmenting by maximum between-class variance of gray-thresholding. The highlights of new algorithm lie in two aspects: the denosing by substacting a background image and the targets’ detection. In virtue of the denosing method, such stable noise signs as stripes and maculae deriving from photography processes are removed effectively from the sand particles’ image. As the most an important precedure of the improved algorithm, the targets’ detection is able to distinguish effectively those dark sand particles from background in differential units by selection of appropriate variance of gray (3.5), but also to reduce fasle information that backgrounds could be recognized wrongly as particles. It seems that image segmented with the targets’ detection shows clearly more sand particles in contrast to image without the targets’ detection which is blurred with lots of stochastic noises. Based on horizontal and vertical coordinates of all sand particle in the study image recorded manually, such parameters as the number of sand particles identified correctly (N_ie), recall rate (R_c) (refers to ratio of the extracted automatically number of sand particles to the number of real particles (N_r)) and the precision (P_r refers to ratio of N_ie to N_r) were used to evaluate the algorithm. It shows that N_ie, R_c and P_r is 461, 71%和86% respectively, compared to 85, 13% and 82% by the traditional algorithm. The new algorithm is better than the traditional one. Nevertheless, it should be perfected in future through many new ways.