Super resolution algorithms are necessary for improving the quality of low resolution video sequences from surveillance cameras. These algorithms have two main problems: first, they hardly can improve the quality of their inputs by factors bigger than two. Second, applying them to real video sequences usually produces unstable and noisy output. The proposed system in this paper deals with these two problems. The latter, which is due to the unavoidable registration errors of video sequences, is dealt with by using a face quality assessment technique. A combination of different types of super resolution algorithms in a hybrid system is used to cope with the former. The system is tested using real world videos from uncontrolled environments and the results are promising.