Applying Kinect sensorsare increasing in recent years due to low price and wide applications. This device can estimate the posture of the human body by utilization of skeletal data and without using any marker. Obstruction of the human body by other objects and rapid movement in front of the Kinect are the main problems in estimating the position of the joints. Two steps are considered in this study to solve the existing challenge.In the first step, a solution based on measurement models has been proposed to determine the reliability of joint position that extracted from Kinect, which is considered as an effective feature associated with the joint position in the Max-Margin classifier. Then, based on the reliability of each joint, a decision is made and the missing joints are identified. Finally, the joints are accredited using human body segmentation algorithms based on a deep learning network. The results show that selection of appropriate features in the first step to verify consecutive frames compared to existing approaches,has a significant improvement in the accuracy of the classification. The second phase also has a supreme impact on the accuracy of the methods that use the Kinect sensor skeletal data as input features by applying validation.