Abstract: Fitness running is a common exercise in sports. Due to the improper running form, it is not only
hard to reach the ideal fitness result, but also easy to harm the body seriously, especially the knee. To
address the problems of lacking scientific guidance and monitoring in current running systems, this paper
proposed a virtual running exercise system based on Kinect. In the action recognition, the dimension of
running data collected from Kinect is reduced according to the angle characteristics and the relative
position of human joints, and running action units are divided from motion features set. Then all kinds of
running actions are recognized by matching with curve characteristics of action units and classification
models. In addition, scientific guidance about their running poses is given to users while exercise. In the
motion modeling, we analyze the relationship between the heart rate and running intensity, and build the
fitness running model and the individualized scientific running exercise scheme, making the heart rate
always in the safe and effective range. The experiment results show that our system is practical and
feasible to provide a scientific method for safe and effective fitness running to users.

Key words: somatosensory interaction, fitness running, action recognition, motion modeling