The document summarizes CMU's submissions to the TRECVID 2011 Surveillance Event Detection task. It describes their framework using MoSIFT features, spatial bags of words, and methods like cascade SVMs and bagging random forests to handle imbalanced data. Their best runs achieved the lowest minimum detection cost rate for 3 out of 6 events compared to other teams, with improvements over their 2010 results. Event-specific window sizes and classifiers like random forests produced better results than last year.