TJU-TJUT@TRECVID 2015: Surveillance Event Detection

In this paper, we present an event detection system evaluated in 2015 TRECVID-SED. The system consists of two parts: automatic event detection (retrospective) and interactive event detection with human in the loop (interactive). For the retrospective part, an extended spatio-temporal features, MoSIFT, is extracted, and two types of temporal divisions (annotation partition and sliding window) are employed as the detection unit. BoW is used to encode low-level features as the representation of each video segment. In order to deal with the highly imbalanced nature of surveillance data, the system performs detections using the proposed Horizontal SVMs algorithm according to each specific event and decision-level post processing is used to combine multiple detection scores. For the interactive part, we designed and developed an interactive visual analytics system, which can enable effective rank detection results with score relations and utilize user feedbacks to improve surveillance event detection.