Comparison of methods to determine pedestrian shoulder orientation from stereo recordings

Abstract

The orientation of pedestrians is used for analysis of space usage, social groups or orientation dynamics in crowds. For this purpose, it is desirable to measure the orientation for each pedestrian in the crowd. This work compares geometric and machine learning methods to determine the shoulder orientation. We make use of stereo camera systems to enable the estimation without the need for intrusive measurement equipment.
The machine learning methods consistently outperform the geometric methods. Especially in higher densities. However, high-density and low movement datasets are still challenging for machine learning techniques with automated labels. In laboratory experiments, coloured shoulder markers can be a cost-effective alternative method.