How to describe face sequences for fast person recognition

The evaluation of video material for forensic purposes is a time intensive and complex work. A common task is to identify or find persons in video footage. Computer vision based methods can help to reduce the manual effort. However, video databases in forensic applications are often rather large. This poses harsh requirements with respect to the processing speed of any automated recognition approach. Specifically, searching for persons needs to be much faster than real time. An analysis and evaluation of existing face recognition techniques is performed with respect to this requirement. Based on this result a promising approach is presented. The key concept is to use a cascade of the existing techniques and combine them in a way that the advantages of each one are used. This results in a significant speedup in processing time and additionally in a slight improvement in the recognition performance. Using this approach promises to help at the forensic search in video footage.