Date Published: June 6, 2019
Publisher: Public Library of Science
Author(s): Claire Chambers, Gaiqing Kong, Kunlin Wei, Konrad Kording, Markus Lappe.
Marker-less video-based pose estimation promises to allow us to do movement science on existing video databases. We revisited the old question of how people synchronize their walking using real world data. We thus applied pose estimation to 348 video segments extracted from YouTube videos of people walking in cities. As in previous, more constrained, research, we find a tendency for pairs of people to walk in phase or in anti-phase with each other. Large video databases, along with pose-estimation algorithms, promise answers to many movement questions without experimentally acquiring new data.
To successfully navigate any environment, a walker must adapt to the surface they walk on, avoid obstacles, change speed, and plan movements according to their goals [1,2]. Processes that impact walking behavior range from peripheral processes to high-level decision-making processes [3,4]. Walking, like all movement, often takes place in the context of other people. Humans, when walking, must often generate their actions according to the movements of people around them.
We asked if people synchronize their walking when they walk side-by-side in naturalistic settings. To do so, we analyzed videos found on YouTube. Within YouTube videos, we searched for video segments with pairs of people walking. From video segments, we extracted the pose of each member of the walking pair. In order to examine walker synchronization, we analyzed displacement between left and right ankles for each member of the pair. Based on the displacement signal, we extracted the walking frequency and mean relative phase for each pair to examine walker synchronization.
We asked how people synchronize their movements when they walk side-by-side, by analyzing pose estimates extracted from online videos. We searched for videos on youtube.com using the search term ’walking in’ followed by the names of major cities. We analyzed the relative phase and walking frequency computed from the vertical displacement between the ankles in each video. We validated our analysis from pose estimates through comparison with ground-truth data and found reasonable agreement between estimates of mean relative phase and walking frequency. The mean frequency of walkers was 1.85 Hz, close to the average of 2 Hz reported in the literature . We found that the distribution of relative phase across video segments contained prominent peaks near 0 and π, which reflects a tendency for pairs of walkers to walk in phase or in anti-phase with each other. Using pose estimation applied to online videos, we confirmed findings from more constrained laboratory experiments on a larger sample and in more naturalistic conditions. We have thus shown that a pose estimation algorithm applied to 2D videos can be used to address questions about naturalistic movement.