There has been a dramatic increase in video data on the Internet in recent years. The abundant video data result in the request for content-based video retrieval (CBVR), which uses efficient techniques to index, retrieve or browse these data. For this purpose, we combine color details derived from dynamic color histogram and texture details derived from fast wavelet histogram to select key frames. Then, the number of key frames is determined by optimized k-means algorithm. Finally, we remove the redundant key frames and get the final key frames by way of mutual information. The effective gradual shot transition is obtained by means of variance curve of dynamic color histogram. Three test videos including movie, sport and lecture are adopted to evaluate the proposed method on performance of key frame extraction and gradual shot detection.