MMM-TJU at TRECVID 2010

Surveillance Event Detection Semantic event detection in the huge amount of surveillance video in both retrospective and real-time styles is essential to a variety of higher-level applications in the public security. In TRECVID 2010, to overcome the limitations of the traditional human action analysis method with human detection/tracking and domain knowledge, we evaluate the general framework for multiple human behaviors modeling with the philosophy of bag of spatiotemporal feature (BoSTF). The brief introduction to each run is shown in Table 1. Table 1 SED RUNs ActDCR Description SED_Runs ActDCR Descriptions TJUMM_1 6.6931 MoSIFT, SVM with 2 χ kernel, 0.50 (threshold) TJUMM_2 5.2067 MoSIFT, SVM with 2 χ kernel, 0.60 (threshold) TJUMM_3 4.4773 MoSIFT, SVM with 2 χ kernel, 0.65 (threshold) TJUMM_4 3.7913 MoSIFT, SVM with 2 χ kernel, 0.70 (threshold) TJUMM_5 3.1070 MoSIFT, SVM with 2 χ kernel, 0.75 (threshold) TJUMM_6 2.4753 MoSIFT, SVM with 2 χ kernel, 0.80 (threshold) TJUMM_7 1.9196 MoSIFT, SVM with 2 χ kernel, 0.85 (threshold) TJUMM_8 1.4527 MoSIFT, SVM with 2 χ kernel, 0.90 (threshold) Semantic Indexing Semantics indexing is extremely helpful for automatic semantic discovery and annotation. In TRECVID 2010, we mainly evaluate three kinds of features, two global features (grid-based color moments and texture feature) , and one local feature (Scale-Invariant Feature Transform, SIFT) for semantic modeling. The cascade Support Vector Machine (SVM) is implemented for each concept model leaning in two ways. First, for each concept three kinds of classifiers are learned with individuals. Second, the decision of three experts above are fused with average fusion algorithm to take advantage of the superiority of individuals. Therefore, we obtained four runs for evaluation of the High Level Feature Extraction test in TRECVID 2010. The brief introduction to each run is shown in Table 2. Table 2 SIN RUNs infMAP Description SIN_Runs InfMAP Descriptions L_A_MMM-TJU1_1 0.0156 GCM Feature and Cascade SVM L_A_MMM-TJU2_2 0.0052 Texture Feature and Cascade SVM L_A_MMM-TJU3_3 0.0238 SIFT Feature and Cascade SVM L_A_MMM-TJU4_4 0.0267 Fusing All Results TRECVID 2010 Surveillance Event Detection by MMM-TJU An-An Liu 1 , Zan Gao 1,2 1 School of Electronic Information Engineering, Tianjin University, Tianjin 300072 P.R. China 2 School of Information and Telecommunication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876, P.R. China