Abstract

     Most developed countries are facing important demographic issues related to ageing populations. Maintaining elders at home while ensuring their safety and well-being often constitutes the main goal of these countries. An interesting solution to this challenge is to develop a smart home, able to monitor the routines of the resident, to recognize the on-going activities, and to provide support when required. In the literature, most works focus on monitoring high-level behaviors such as eating, sleeping, etc. However, to provide live guidance, the system needs to have a far more detailed recognition process able to identify the specific steps of the on-going task and the erroneous executions. In this paper, we propose an algorithmic approach for hand gesture recognition designed to be used as the core component of a fine-grained activity recognition model. It is based on the processing of inertial data collected from a wristband equipped with triaxial accelerometer and gyroscope, and machine learning techniques. A simple set of gestures for cooking activities as been defined, enabling characterizing high-level cooking tasks. To do that, we constructed a labelled dataset of atomic gestures performed by participants that we made available to the scientific community. We obtained an average accuracy of 0.83 in recognizing the gestures with the leave-one-subject-out strategy.

Introduction

     Western countries are currently experiencing an unprecedented demographic crisis linked to the accelerated ageing of their population [1]. This reality is worsened by a problem of a general labor shortage [2], particularly with regard to qualified personnel in the medical field, and more specifically for home care dedicated to people with loss of autonomy. Seniors suffering from a loss of autonomy wish to remain at home [3]. Governments are pushing for this for both social and economic reasons. In fact, keeping elders at home longer is clearly desirable because it contributes improving the quality of life, which corresponds to societal values: people should live as normal a life as possible without segregation in retirement homes. However, keeping seniors with loss of autonomy at home involves many risks that need to be controlled. The physical environment of residences must therefore be adapted, or even increased with the help of technology and artificial intelligence, in order to meet elders’ needs, to compensate for cognitive and physical disabilities, to ensure safety and to adequately support residents in carrying out their Daily Living Activities (ADL) [4].

Conclusion and Future Works

     In this paper, we proposed a model for hand gesture recognition designed to be used as the core component of a fine-grained activity recognition system. It relies on a wristband with a triaxial accelerometer and a triaxial gyroscope. We defined a set of gestures for cooking activities, which allows characterizing high-level cooking tasks as a set of simple gestures. One of the main contributions of our work is the construction of a carefully labelled dataset of 13 gestures gathered from 21 participants using a rigorous experiment protocol. This dataset has been made available to the scientific community. Each participant performed each gesture a minimum of 20 times. The system showed good results with an average accuracy of 0.83 in recognizing the targeted atomic gestures with the LOSO strategy. Also, these results are promising for developing a complete fine-grained cooking activities recognition model. In future works, we plan to add a mechanism that prioritizes certain types of gestures in case of multiple simultaneous recognition. For instance, a fast gesture of one second, or two, can be recognized in the same three seconds window in which we could also have recognized a slow gesture. In that case, we need to discriminate. Moreover, the atomic gestures recognition system could be enhanced with ambient sensors. Finally, we also plan to extend the actual system using RFID sensors to approximate the 2D position of the hand in the room versus the actual position of different detectable objects.