Classification of Body Posture on TikTok Videos Using a Representation Based on Angles

Abstract

Representing complex phenomena, such that coming from video sources, in order to build models for posture detection involves challenges such as: avoiding loss of relevant information, rely on that the representation fits the expectative, and avoid cost for specialized hardware [1, 4, 5, 8].

According to Patel et al. [7] detecting corporal movement is gaining strength for tasks such as: personal health care, environmental awareness, human-computer-interaction and surveillance systems.

On the other hand, as we previously said, acquiring information based on corporal movement to feed the algorithms that build the models can be costly in time, computational resources, human participation, as well as the variety of representative movements to be studied. However, considering crowdsourcing coming from social networks which are focused in creating contents such as TikTok has facilitated collecting data from this nature. In particular, the contents in this social network are focused on corporal movements such as dancing, walking, speaking in sign language etc. which can be used to integrate large and diverse data sets.

The trending in the platform, which are mainly oriented to dancing, make it possible to collect postural data sets which are arranged into similar observations. Nevertheless, there exists a challenge to analyze these sources due to the diversity to record the videos, for example we can find different resolution, backgrounds, source light, clothes, spatial position from the individual, etc. In this sense, preprocessing video is one of the first tasks to tackle noisy or diverse information and provide reliable information in the subsequent stage of the analysis.

Therefore, in this article we propose a novel technique for extracting information that can be used to generalize postural information from the human body in TikTok videos without being affected by the differences in their capture methodology. To build a data set with the initial features from the videos we used the algorithm proposed by Cao et al. [2]. In our case, the novel technique is based on the representation of relationships between the corporal features which can form triangles in their significant positions within the posture. Due to such representation it is possible to search for equivalence in angles through different data sets. In that way, we reduce the number of variables used to find similarities, enabling the possibility to create generalized models with lower costs. In this case we used a Naïve Bayes Network to verify the level of generalization and performance.