Ph.D. Tezi Görüntüleme | |||||||||||||||||||||
|
|
||||||||||||||||||||
| Summary: Video processing work has been worked since 1980s. Since that time, human activity recognition has become one of the most challenging tasks in the field of computer vision. Despite the studies that belong to the subject, many problems related to action recognition have not been solved yet. In recent years, with the emergence of the Microsoft Kinect sensor and the resurgence of deep learning methods, is provided cost efficient and reliable 3D human skeleton. In this thesis, a bag of pose method which uses 3D skeletal data for the human action recognition have been proposed. In this study each action is represented by a set of predefined Spatio temporal key poses. The definition of temporal spatial descriptors to represent 3D poses is the main contribute of the study. The pose descriptors are consist of three parts concatenation. The first part is the normalized positions of 3D skeleton joints. The second is the displacement of the same joints of the poses over a predetermined time offset and the third part is the displacement vectors that obtained from the joints of the current and the previous skeleton. The Key poses has obtained by applying k means clustering method on all of training poses. Later every action has been converted to a sequence of key poses and key poses histograms has obtained. In the last stage ELM was used as classifier. For the evaluation of the proposed method, have been used five popular action data sets that have 3D skeleton. Achieved state of the art results on three of the datasets and competitive results on the other two datasets compared to the other methods.Key Words: Skeleton based 3D action recognition, Bag of words, Key poses, Extreme learning machine, RGB D. | |||||||||||||||||||||