CRIM Notebook Paper - TRECVID 2011 Surveillance Event Detection

Approach we have tested in each of your submitted runs. For the “Object Put” event, we followed a dual foreground segmentation approach where the output difference between a short term and a long term model is used for triggering potential alerts. For Pointing, Embrace, CellToEar and PersonRuns, we applied the learning of compound spatio-temporal features based on a data mining method. Relative contribution of each component of our approach. Our system is based on an action recognition approach which is mining spatio-temporal corners in order to detect configurations, called Compound Features, typical of an action of interest. The final detection is based on blobs around local frame-to-frame changes that are containing enough relevant compound features. What we learned about runs/approaches and the research question(s) that motivated them. Overall, performances have improved from last year especially for PersonRuns. In addition, the training for PersonRuns was based on a standard action recognition dataset (KTH) independent of the TrecVid dataset which indicates that our implementation is behaving as expected. For Pointing, Embrace and CellToEar, results are not satisfying yet and the main reason is probably due to the fact that the training dataset derived from the development videos presents a large variability, is too noisy and too small in size in order to produce good rules. Also, given the complexity of the scenes composed of multiple action occurrences, occlusions and complex actions, the direct application of an action recognition method is a challenge. Going forward, performances could be improved if combined with other approaches such as a person tracker and also if the quality of the training set could be improved.