Data Driven Kinematic Modeling of Human Gait for Synthesize Joint Trajectory

Synthesis of reference joint trajectories for the legged robot is a very difficult task due to higher degrees of free-dom. The gait dataset can be used to develop the models which can provide the required references. This paper presents the kine-matic modeling of human gait data, which is used as the reference joint trajectory for a Biped robot, 8 deep learning models are proposed. Gait data-set of 120 subjects are collected at RAMAN Lab, MNIT Jaipur, India using the vision-based methodology. All subjects belong to the 5–60 years age group. Four type of novel mappings, one-to-one (knee-to-knee, hip-to-hip, and ankle-to-ankle), many-to-one (knee+hip+ankle-to-knee/hip/ankle), one-to-many (knee/ankle/hip-to-knee+hip+ankle), and many-to-many (knee+hip+ankle-to-knee+hip+ankle), are also developed. These mapping provides the reference trajectories to biped robot and relationships between the knee/hip/ankle trajectories is also ob-tained. Performance evaluation of developed models is measured by average error, maximum error and root mean square error. Results show that the bidirectional deep learning technique performs better for different mappings. Finally, a discussion is provided for the applicability of developed mapping robots in real biped robots.