To evaluate the work contents of a care worker, this paper proposes a novel lower limb posture estimation method, and a motion recognition method for the automation of care records. In our approach, wearable-embedded sensors consisting of inertial sensors and insole-type plantar pressure sensors were used to recognize the posture of the whole body during care tasks. The upper body posture was calculated from triaxial accelerations, and posture classification of lower limbs can be reached with characteristics extracted from plantar pressure distribution. To achieve accurate posture recognition considering unexpected posture in training phase and compound postures combining multiple postures, a new posture estimation method combining normal and complementary Gaussian mixture network (NACGMN) and posture-fitness function was developed. Using the time-series data of posture information, our proposed work estimation manner utilizing hidden semi-Markov models can recognize performed care works based on the frequency distribution of state transitions. In experiments, we measured and recognized transfer assistance works between a bed and a wheelchair in our laboratory and an actual care facility, and the results demonstrated the effectiveness of the proposed system.
