As a key technology in content-based video indexing and management, the video shot boundary detection has attracted considerable research attention. However, the traditional detection algorithm can only handle a hard cut boundary. No satisfactory results have been achieved on gradual transition until now. A method based on the combination of adaptive threshold and Fourier fitting is presented in this paper to detect shot transition. A non-uniform histogram in HSV color space on each frame is first accumulated to calculate a similarity sequence of videos, depending on which thresholds are generated by the adaptive threshold to detect hard cut transition. Gradual transition has a much more complex transform in its span. After extensive experiments and observation, the study has found that fixed patterns exist in every gradual transition. Candidate transitions are detected by finding segments which show traits of a gradual transition pattern. Gradual transition boundaries of different types are collected to train a set of standard templates, which can be used to judge whether a candidate is a real gradual transition boundary and also, to further determine its transition type. The algorithm is accelerated by exploiting the parallel computation power of GPUs using CUDA. The effectiveness is verified through extensive experiments and is compared with other methods.