國網中心

Video surveillance systems are deployed at many public spaces for Urban security and surveillance purposes such as airports, train stations, and shopping malls. However, it is laborious to analysis and retrieval for specific persons in multiple camera surveillance systems, especially in the cluttered background and appearance variations among multiple cameras. In order to address the problem, this system presents an AI video analysis and retrieval method to extract the human attribute via deep-learning instance-segmentation technology. It uses attributes like clothes color and type to describe a person. The proposed system of person retrieval consists of four steps: (i) using deep-learning instance-segmentation technology to perform pixel-wise person segmentation; (ii) appearance-based attribute features with multi-CNN; (iii) search engine with fundamental attributes matching approach; and (iv) the video summarization technique to produce temporal abstraction of retrieval objects.

National Center for High-performance Computing (NCHC) cooperates with the Massachusetts Institute of Technology (MIT) and Argonne National Laboratory to promote the application of smart metropolis with high-speed computers and AI Deep Learning to solve congestion problems in main cities. The project further extends to cooperating with National Chung Hsing University (NCHU) and TungHai University (THU), we successfully applied the AOT urban monitoring system and sensor analysis to environmental monitoring. In practice, the system was applied to the cooperation of the Taichung City Government and the Hsinchu City Government to promote the cloud IoT of smart cities to solve the issues of monitoring data convergence and big data application in urban data governance. The results of the first phase of this project including the construction and application of the AOT monitoring system, real-time environmental monitoring in Taichung City, and traffic congestion analysis in the Central Taiwan Science Park (CTSP). The analysis of the traffic congestion in the CTSP is based on the intersection information of CTSP as the experimental environment and using MIT's traffic flow simulation system, and the traffic flows analysis system jointly developed by NCHC and Taiwan Yuan Ze University. These two systems automatically analyze the traffic flow at intersections and learn how to optimize traffic signal allocation to alleviate traffic congestion during peak time.

We intend to present a rainfall-inundation simulation database which comprises a big number of simulated rainfall-induced inundations by collaborating the storm generator with 2D hydraulic numerical model successfully adapted in NCHC TWCC platform. In detail, the storm generator mainly reproduces gridded rainfall data with high resolution in time and space (i.e. quantitative precipitation estimaton, QPE) by means of Monte Carlo simulation method for the non-normal correlated multi-variates with the gridded rainfall characteristics under the consideration of correlation structure in time and space. After that, the simulated gridded rainstorms are imported into the 2D hydraulic numerical model (SOBEK) to carry out the inundation simulation. Eventually, a big number of rainfall-induced simulations can be obtained so as to create a databased regarding the rainfall-inundation simulation. Accordingly, through TWCC platform, the resulting rainfall-inundation-simulation database can be applied in training a variety of AI models for the flooding forecasts, including the artificial neural network (ANN), the long short-term memory (LSTM), convolutional Neural Network (CNN) and support vector machine (SVM) as shown in the webpage. The relevant results from training AI models using the proposed rainfall-inundation simulations database indicate that the aforementioned AI models are well-established ones which can provide information on the flooding forecasts.