Microsoft Kinect can be used for computationally inexpensive acquisition of skeleton tracking in real time. For human activity recognition, it appears to provide an opportunity for researchers to achieve good performance at low cost. However, two issues still remain. Firstly, the Kinect skeleton tracker often captures unnatural skeleton poses, such as discontinuous and vibrated motions, in the presence of self-occlusion. Secondly, there is still a requirement for anyone wishing to understand human behavior to develop high-level features instead of making direct use of a 3D skeleton pose. To this end, we propose a method that is composed of two parts. The first part is to improve the Kinect skeleton under self-occlusion by using deep recurrent neural networks. The second part is to extract features by evaluating the importance of each subsequence of trajectories using a complexity-based measure.