Causality-weighted active learning for abnormal event identification based on the topic model

Abstract

Abnormal event identification in crowded scenes is a fundamental task for video surveillance. However, it is still challenging for most current approaches because of the general insufficiency of labeled data for training, particularly for abnormal data. We propose a novel active-supervised joint topic model for learning activity and training sample collection. First, a multi-class topic model is constructed based on the initial training data. Then the remaining unlabeled data stream is surveyed. The system actively decides whether it can label a new sample by itself or if it has to ask a human annotator. After each query, the current model is incrementally updated. To alleviate class imbalance, causality-weighted method is applied to both likelihood and uncertainty sampling for active learning. Furthermore, a combination of a new measure termed query entropy and the overall classification accuracy is used for assessing the model performance. Experimental results on two real-world traffic videos for abnormal event identification tasks demonstrate the effectiveness of the proposed method.