Detecting human actions in surveillance videos

This notebook paper summarizes Team NEC-UIUC's approaches for TRECVid 2009 Evaluation of Surveillance Event Detection. Our submissions include two types of systems. One system employs the brute force search method to test each space-time location in the video by a binary classifier on whether a specific event occurs. The other system takes advantage of human detection and tracking to avoid the costly brute force search and evaluates the candidate space-time cubes by combining 3D convolutional neural networks (CNN) and SVM classifiers based on bag-ofwords local features to detect the presence of events of interests. Via thorough cross-validation on the development set, we select proper combining weights and thresholds to minimize the detection cost rates (DCR). Our systems achieve good performance on event categories which involve actions of a single person, e.g. CellToEar, ObjectPut, and Pointing.