eCubeLand: An Intelligent Multi-view Video Data Modeling

The extensive use of surveillance systems, particularly those installed in Internet of Things environments, leads to the continuous harvesting of tremendous amounts of video data. The effective analysis and management of these data are challenging tasks for surveillance experts due to unstructured storage and variability. We propose an intelligent modeling framework, offering a convenient representation with indexing for real-world objects and solving complicated computer vision problems, such as anomaly detection and person re-identification. Moreover, our framework generates grids and assigns indexing to visual sensors and real-world entities, allowing efficient information retrieval with better resource allocation. The proposed framework consists of four major modules: 1) mapping, 2) data analysis, 3) information sharing, and 4) data storage. The mapping module is responsible for analyzing the environment, followed by the data analysis module, which detects, analyzes, and indexes the entities. Furthermore, video data from these modules are passed to the information sharing module, which generates alerts in the case of undesirable scenes and broadcasts the meaningful information toward adjacent visual sensors. The final module is used to preserve anomalous data along with the identified person’s information from distributed vision sensors. To validate the proposed framework, we perform experiments on real-world complex tasks, including anomaly detection and person re-identification, showing promising performance on surveillance video data.