With the construction of smart city and the development of Internet of Things (IOT) technology, intelligent video surveillance systems have received increasing attention and application. Meanwhile, as the trend of development has been towards internetworking, high definition and ultra-high definition, the video data is increasing exponentially but with a low effective utilization rate. Video storage, browsing, and content analysis become urgent issues in academia and industry. Visual scene activity analysis and video summarization technology are two important means to achieve massive video intelligent management, which allow us to quickly browse and precisely locate events occurring in videos. Aiming at these two issues, we have studied the following topics: 1. We propose a context learning based method for collective activity classification. A collective activity is shared by all individuals presented in the group, and maybe also interacts with other groups. Potential functions are introduced to describe the compatibility between individual action and appearance feature, the relationship between individuals within the same group, and also the interaction among different groups. A discriminative structure SVM model is proposed to jointly learn these contextual information in a unified framework. Experimental results have demonstrated the superiority of our approach in collective activity classification. 2. Based on the theory of bi-layer sparse topic model (BiSTM), we propose an unsupervised approach for dynamic scene understanding. The input surveillance video is represented by a topic model, where clips without overlapping are treated as the document while low-level visual features are quantized into discrete words. Then motion pattern mining is converted to a problem of learning latent topics. In the BiSTM model, both the topic level sparsity and the document level sparsity guarantee the semantics of motion pattern. In addition, considering the characteristic of extreme imbalance between numerous typical normal activities and few rare abnormalities in surveillance video data, the one-class learning problem is introduced and a discriminative BiSTM is proposed for abnormality detection. Experimental results and comparisons demonstrate the promising performance of the proposed approach. 3. We propose a novel presentation approach to vividly depict the moving process of a specific object in a surveillance video, which aims at effectively summarizing...
修改评论