Abstract
Human-Computer Interfaces have always played a fundamental role in usability and commands’ interpretability of the modern software systems. With the explosion of the Artificial Intelligence concept, such interfaces have begun to fill the gap between the user and the system itself, further evolving in Adaptive User Interfaces (AUI). Meta Interfaces are a further step towards the user, and they aim at supporting the human activities in an ambient interactive space; in such a way, the user can control the surrounding space and interact with it. This work aims at proposing a meta user interface that exploits the Put That There paradigm to enable the user to fast interaction by employing natural language and gestures. The application scenario is a video surveillance control room, in which the speed of actions and reactions is fundamental for urban safety and driver and pedestrian security. The interaction is oriented towards three environments:the first is the control room itself, in which the operator can organize the views of the monitors related to the cameras on site by vocal commands and gestures, as well as conveying the audio on the headset or in the speakers of the room. The second one is related to the control of the video, in order to go back and forth to a particular scene showing specific events, or zoom in/out a particular camera; the third allows the operator to send rescue vehicle in a particular street, in case of need. The gestures data are acquired through a Microsoft Kinect 2 which captures pointing and gestures allowing the user to interact multimodally thus increasing the naturalness of the interaction; the related module maps the movement information to a particular instruction, also supported by vocal commands which enable its execution. Vocal commands are mapped by means of the LUIS (Language Understanding) framework by Microsoft, which helps to yield a fast deploy of the application; furthermore, LUIS guarantees the possibility to extend the dominion related command list so as to constantly improve and update the model. A testbed procedure investigates both the system usability and multimodal recognition performances. Multimodal sentence error rate (intended as the number of incorrectly recognized utterances even for a single item) is around 15%, given by the combination of possible failures both in the ASR and gesture recognition model. However, intent classification performances present, on average across different users, accuracy ranging around 89–92% thus indicating that most of the errors in multimodal sentences lie on the slot filling task. Usability has been evaluated through task completion paradigm (including interaction duration and activity on affordances counts per task), learning curve measurements, a posteriori questionnaires.
