Abnormal Activity Detection based on Dense Spatial-Temporal Features and Improved One-Class Learning
Tam ngoc Nguyen, Ngoc Quoc Ly · 2017
Abnormal activity detection is an important issue in video surveillance. The abnormal activity could be a predictable activity or unpredictable activity. This paper focuses on unpredictable activity detection. Due to unpredictable anomalies, we do not have training data of them, so we could not use the discriminative learning model to detect abnormal activity and normal activity. One class learning method is the generative model and it is suitable to model unpredictable abnormal activities. In this paper, we use fast dense spatial-temporal features within regions of interest points to model normal activities by Support Vector Data Description (SVDD). Besides. we use K-means++ algorithm to cluster normal data then the multi hyperspheres SVDD are constructed separately on clusters instead of only one hypersphere SVDD on multi-distribution data. Experiments on benchmark datasets contain various situations with human crowds, overlapping between individual subjects and low resolution. The experiments show that our approach could outperform some state of the art methods on the Ped2 dataset.