Ward-CMu@trecvid 2015 surveillance event detection

We present a retrospective system for event detection in surveillance videos automatically, which is built on the Gatwick development data. It is an enhanced version of the retrospective system in [1]. The changes come from four aspects. First, dense trajectory [2] and improved dense trajectory [3] are used together in the proposed system. Second, the PCA features used in Gaussian Mixture Model are changed to whiten-PCA features. Third, we implement a learning-based probability function for LIBLINEAR [4]. Forth, instead of averaging all the detection scores we have, we select two kinds of them to get better results. We think all the changes are beneficial to the final submission ‘WARD-CMU p-fusion_1’ which wins 4 events in all 7 events. Specifically, it is worth noting that the PersonRuns sets a new record in recent years’ SED competitions. Through the results in our internal evaluation, we think dense trajectory and improved dense trajectory are complementary for event detection in a complex surveillance environment. We also notice that current system is bad at detecting the short events like CellToEar, ObjectPut and Pointing. We are considering to introduce other methods such as pose estimation and pedestrian detection to enhance current system in the future.