Gesture-based incident reporting through smart watches

Abstract

With the capabilities and autonomy of smart watches constantly increasing, there is the need for introducing applications which can exploit their full potential, establishing the use of smart watches in our daily routine. The field of personal safety and security provides an excellent basis over which applications can be developed, enabling the use of wrist worn devices as tools for easy, discreet and efficient reporting of incidents or suspicious behavior. However, current practices in report creation using smart watches rely on methods and interfaces, such as taking pictures and writing text, without taking into account gesture-based input. In this paper, we present the design a smart watch-based approach, which utilizes a Deep Learning model, to recognize specific user gestures that could result in reporting hazardous situations and could alert the authorities for assistance. We evaluated the performance of the model by training it to distinguish 5 predefined gestures from a set of random gestures performed by 9 subjects wearing a smart watch on their dominant arm. Our Deep Learning model surpasses the performance of conventional classifiers that rely on hand-crafted features and produced gesture recognition accuracies above 98% in 26,061 motion signal samples by fusing the automatically extracted features of 3-axial accelerometer signals. We conclude by discussing the related issues we have encountered by using the proposed application in real-time use and providing future directions.