A Feasibility Study for Health and Life-Threatening Conditions Recognition via Sensor Fusion Approach

The paper presents a feasibility study for health and life-threatening conditions such as falls or postural balance abnormalities recognition via sensor fusion approach. As a sensor fusion system we used the Intel RealSense Depth Camera D435 that combines standard video and depth cameras. However, the possibility of recognition of different types of movement patterns especially falls with RGB-D cameras has been studied for more than a decade now, the usage of such methods to detect life and health-threatening conditions which may be manifested by impaired postural balance has not been paid enough attention yet. The proposed approach was validated on an experimental dataset collected for two experimental surroundings with participation of 13 volunteers (9 females and 4 males, age: 19–37 years). Each subject performed 11 different types of motion patterns including 2 patterns specific for life and health threatening situations. We used the Caffe neural network to find key point coordinated and form the feature vector than was used for training models. The proposed classifier for video sensor data classification showed 81 % accuracy and Cohen’s kappa of 78 %, while the usage of sensor fusion approach reduces the number of errors, and increase the classification accuracy and Cohen’s kappa by 7 and 8 %, respectively. Nevertheless, the economic feasibility of such a solution requires an additional assessment.