Gait Estimation of Occluded ROIs Using Interpolation Techniques and Testing Their Performance in Speed Variation in Gait

In human Gait analysis, one of the major obstacles is occlusion, occlusion is a scenario where a specific portion of the human body part gets hidden under some obstacles. In a multi-person gait scenario, a subject’s body part gets occluded by another subject’s body parts. One of the main concerns in multi-person gait is, how to estimate the occluded frames keypoint’s value effectively. In this paper, 3 interpolation techniques Cubic spline, Piece Wise Cubic Hermite (PCH), and AKIMA 1-D interpolator have been studied on two different Gait databases. Ten keypoints that have a very high contribution to Gait Cycle have been considered. These Points are the Shoulder joint, Elbow, Wrist, knee, and Ankle on both sides of the body. While testing on the SMVDU dataset, feature extraction of the Keypoints Coordinates has been performed by using the CMU-OpenPose library, as a next step occlusion is detected, if it exists for all the keypoints, using a novel occlusion detection algorithm. In the third step above mentioned interpolation techniques are applied to estimate the occluded Region of Interest (ROI). Finally, quantitative assessment techniques MSE, RMSE, MAE, and MAPE have been used and it is found that the AKIMA-1D interpolator has the lowest estimation error rate in the estimation of occluded ROIs for most of keypoints. To Test the performance of the interpolation techniques while estimating occlusion under a different mode of speed, ANDERSON.et.al designed Kinect dataset has been considered, which has multiple samples of the same subject walking at a different speed. In this database also AKIMA-1D interpolator has given the lowest estimation error for most of the key points.