Surveillance cameras are increasingly being installed to detect public anomalies and prevent crimes. However, such additional surveillance cameras require a larger workforce to monitor videos, inducing heavy costs. Various methods have been proposed to automate anomaly detection in surveillance videos, including deep-learning-based methods. Unfortunately, state-of-the-art methods based on deep learning have not yet achieved a high level of anomaly localization. Our previous study proposed a crowdsourcing-based anomaly detection method which utilizes human cognitive abilities on demand. In an evaluation experiment using shoplifting videos, our method demonstrated high accuracy. In this study, we extended the application of our crowdsourcing-based method to anomaly detection, which is a broader application area. Because a method that relies entirely on crowdworkers is costly, we propose a new crowd-aided anomaly detection method that supports deep-learning-based methods by employing a crowdsourcing-based method, which provides high anomaly localization ability at a low cost.

Amazon Web Services (AWS) provides its excess computing resources as spot instances. Although spot instances are cheaper than on-demand instances, they may be removed by AWS when demand increases for computing resources. Consumers need to judiciously select and prioritize those resources whose operations must avoid being interrupted. Although recently a spot placement score (SPS) is being provided as a measure of spot instance availability, few researches are using SPS. In this study, we propose an instance selection algorithm using it to select spot instances that reduce interruptions.

The recognition accuracy of cloud speech recognition systems has been improving year after year. However, high recognition accuracy is not ensured when multiple speakers are involved. In many situations with multiple speakers, such as in meetings and conversations, voices overlap; while humans can easily distinguish these voices, existing systems have much more difficulty. In this study, we propose and examine a crowdsourcing-based method for identifying and transcribing multispeaker conversations. Crowdsourcing allows users to request work from an unspecified number of workers via the Internet. We demonstrate the usefulness of the proposed system by comparing its accuracy with existing systems and discuss ways to improve the results.