The disclosed method calculates video time density functions based on inter-frame mutual information or other similarity measures. The method includes acquiring a video sequence from memory, computing mutual information or other similarity measures between two consecutive frames of the video sequence, and constructing a video temporal density function based on the mutual information or similarity measures. The method enables fast navigation of the video sequence by performing a temporal quantization of the video time density function to determine a set of quanta. The video sequence can be navigated using the nearest neighbor video frames to the set of quanta. The method enables thumbnail extraction of a video sequence using statistical modeling by performing a temporal quantization of the video time density function to segment video sequence in time domain and using mixture (such as Gaussian mixture and ICA mixture) vector quantization to find the sample frames for each temporal segment.