Strategies to Improve Activity Recognition Based on Skeletal Tracking: Applying Restrictions Regarding Body Parts and Similarity Boundaries

Abstract

This paper aims to improve activity recognition systems based on skeletal tracking through the study of two different strategies (and its combination): (a) specialized body parts analysis and (b) stricter restrictions for the most easily detectable activities. The study was performed using the Extended Body-Angles Algorithm, which is able to analyze activities using only a single key sample. This system allows to select, for each considered activity, which are its relevant joints, which makes it possible to monitor the body of the user selecting only a subset of the same. But this feature of the system has both advantages and disadvantages. As a consequence, in the past we had some difficulties with the recognition of activities that only have a small subset of the joints of the body as relevant. The goal of this work, therefore, is to analyze the effect produced by the application of several strategies on the results of an activity recognition system based on skeletal tracking joint oriented devices. Strategies that we applied with the purpose of improve the recognition rates of the activities with a small subset of relevant joints. Through the results of this work, we aim to give the scientific community some first indications about which considered strategy is better.

Suitability of the Kinect Sensor and Leap Motion Controller—A Literature Review

“…To increase the recognition with skeletal tracking the Extended Body-Angles Algorithm (E-BA-A) [ 116 ] were used in the study of Gutiérrez-López-Franca et al in 2018 [ 117 ]. During the study they found out that the number of the used joints in the body during measurements affects the number of errors.…”

Graph Cut-Based Human Body Segmentation in Color Images Using Skeleton Information from the Depth Sensor

“…In our application of human body segmentation, the skeleton data can be one of the best candidates for the shape prior and the skeleton information can be readily obtained in depth images. In particular, we use a Microsft Kinect v2 device with SDK, which has been extensively used in various applications [20,21,22]. The Kinect SDK extracts the foreground region for extracting the skeleton in real-time using the random forest classifier [11].…”