Estimating human body orientation from image depth data and its implementation

This paper proposes a human body orientation estimation method using the Kinect camera depth data. The input of our system consists of three one-dimensional distance-based signals which reflect the body's surface contours of the human upper body portion, i.e., the upper chest, upper abdomen, and lower abdomen. Such signals are then normalized using their distances to achieve the same amount of the lower parts. All normalized signals are concatenated to provide a mix of contour features. We used Support Vector Regression (SVR) to classify the feature and Kalman Filter to estimate the continuous orientations instead of using discrete orientations. We also extend our work by adding human motion direction to the robust estimate of human body orientation when walking. We conducted two evaluation schemes, i.e., body orientation at static position and body orientation when moving. The experimental results show that our system achieves impressive results by achieving mean average of angle error (MAAE) of 0.097 degrees and 5.82 degrees for estimating body continuous orientation at static position and estimating body continuous orientation when moving, respectively. Therefore, it is very promising to be applied in real implementations.