Fitting distal limb segments for accurate skeletonization in human action recognition

This paper presents a novel method for detecting distal limb segments for accurate skeletonization of human limbs in visual data for human action recognition. After background subtraction, a medial axis transform algorithm is applied to the body silhouette to detect the torso and the limbs. Then, a nine-segment skeleton model is fitted to the medial axis using a line fitting algorithm. The fitting is performed independently for each limb to speed-up the fitting process, avoiding the combinatorial complexity problems. The nine-segment skeleton model is used to provide precise endpoints of the distal segments of each limb which are reduced to centroids for efficient action representation. We believe that the distal limb segments such as forearms and shins provide sufficient and compact information for human action recognition. Each limb centroid is described by its angle, with respect to the vertical body axis, to create a six-element descriptor vector to represent the position of the torso and five angles for limb segments. The nine-segment skeleton model is detected and tracked without any manual initialization. A Gaussian Mixture Model is used to represent action descriptors for several human actions. Then, maximum log-likelihood criterion is utilized to classify actions. To evaluate our approach, we used three action datasets with different resolution and the results are compared with other approaches. As a result, a maximum average recognition rate of 98% is achieved for high resolution dataset and a minimum 90% for low resolution dataset.