Automatic segmentation of therapeutic exercises motion data based on a predictive event approach

This paper presents a predictive event approach which is applied for automatic segmentation of therapeutic exercise sequences. The main purpose of our study is to enhance the efficiency of movement analysis in patients with neurological disorders. When dealing with analysis of rehabilitation exercises it is often required to pre-process the collected data. In order to examine the movements and to carry out the relevant measurements and determine the parameters of interest, each movement must be segmented from a given sequence and analyzed separately. More generally, in case of gesture recognition tasks, after gesture acquisition, segmentation is the following and central step, which can have a huge impact on the classification rate of the gesture recognition system. Gesture segmentation and recognition systems have significant applications in many different fields such as virtual and augmented reality [1], industrial process control [2], physical rehabilitation [3], human-robot interaction [4], computer games [5] etc. Due to this, gesture segmentation is an active research topic and challenging scientific problem, fairly present in the latest research. The predictive event approach is based on a principle of detecting an event when sensor data depart significantly from an adaptive model-based predictor. We have used portable, low-cost Kinect device with marker-free based technique. During exercise execution, the skeleton is continuously detected and 3D positions of characteristic human joints are collected for each frame (Fig. 2). From the original data set, which consists of all collected joints motion data, we have extracted the ones from interest for upper body movement therapeutic exercises (hand and elbow joint). Trajectories of selected joints are modelled as Gaussian processes. Based on this data set, a Gaussian process based predictor is adapted and used to detect significant changes in the exercise sequences. The results over the formed dataset are compared with commonly used technique and illustrate the superiority of our approach.