A hybrid framework for automatic joint detection of human poses in depth frames. (May 2018)

• Record Type: Journal Article
• Authors: Kong, Longbo; Yuan, Xiaohui; Maharjan, Amar Man
• Is Part Of: Pattern recognition. Volume 77(2018:May)
• Page Start: 216
• Page End: 225
• Publication Date: 2018-05
• Subjects: Human pose detection -- Joint detection -- Human body model -- Geodesic features; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.12.020
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A hybrid framework to detect joints from depth videos of human body. Geometric features based on the geodesic distance are used to construct a skeleton with joints and extreme points. A data-driven method is developed to identify joint candidates with weak constraints. Improved performance in terms of accuracy in comparison with the state of the art methods. Abstract: Human pose detection has been an active research topic, and many studies have been done to address different problems in the topic. However, very few methods are proposed to detect joints in the human body. In this paper, we proposed a novel hybrid framework to detect joints automatically by using depth camera. In the proposed method, joints are categorized into two classes: implicit joints and dominant joints. Implicit joints are the joints on the torso, such as neck and shoulders. Dominant joints include elbows and knees. In the hybrid framework we proposed, a loose skeleton model is used to locate implicit joints, and data-driven method is applied to detect dominant joints. The highlight of the proposed work is that geodesic features of the human body are used to build the skeleton model and detect joints. To evaluate our work, experiments are conducted on the dataset recorded by a Microsoft Kinect and compared with state-of-art methods. The results demonstrate that the proposed work can deliver stable and accurate detection results of joints.