Human Body Joints Estimation for Clinical Jumping Analysis

This paper presents an effective approach to estimate human body joints from a monocular video captured with a handheld camera for clinical jumping analysis. In this framework, the video frames are classified into color blobs and represented by region adjacency graphs (RAGs). Then, the corresponding body parts in the current frame are extracted and tracked based on the labels of RAG nodes in the previous frame using a semantic graph growing method. Initially, each RAG node of the current frame is associated with its most similar RAG node of the previous frame. Then, in order to reduce the mismatches in the initial association, the skeleton of legs is constructed to find the correct leg parts. In addition, a loose stick figure model is used to disambiguate the misassignment by enforcing geometric constraints defined between consecutive frames. Finally, the joint positions are estimated and smoothed using a priori knowledge of the jumping process. Experimental results demonstrate the effectiveness and robustness of our algorithm.