Invisible Walls | how pedestrians navigate around social interactions

Jack Terwilliger, Julia Di Silvestri, Seika Murase, Anne Elizabeth Clark White, Federico Rossano

University of California, San Diego

Fig. 1 | A diagram of our field experiments. Actors (shown in orange) stood in busy pathways while displaying 5 distinct signs of interactional involvement (see depictions of Variables 1-5). Over various conditions, we observed whether pedestrians (shown in blue) walked between or around our actors and we measured pedestrians’ walking trajectories.

Abstract

Forecasting pedestrian behavior remains an open problem in robotics, partly due to emergent physical dynamics resembling complex systems. We argue that social dynamics also play a critical role in pedestrian navigation. People interacting (e.g. conversing) on a path impose variable proxemic norms on pedestrians – tacit social rules that regulate the use of physical space. In 4 field experiments with 4,911 participants, we show that pedestrians are acutely aware of others’ social interactions in that they create an invisible wall between interactants that pedestrians try to avoid breaching. Breaching depends on how interactants display involvement in social interactions (body orientation, gaze, and proximity) and whether preceding pedestrians had breached. These results demonstrate how physical mobility in public places depends on pedestrians’ social computations.

Vid. 1 | A demonstration of proxemic norms and their effects on pedestrian traffic flow.

Vid. 2 | Experiment 1 conditions.

Vid. 3 | Experiment 2 conditions & qualitative tracking results.

Vid. 4 | Experiment 3 conditions.

Vid. 5 | Experiment 4 conditions and qualitative tracking results.

Vid. 6 | Qualitative collective breaching results.

Explorable 1 | Interactive exploration of experiment 2 trajectory data from 2023-12-07.

Explore more trajectory data by date

Fig. 2 | A. Video frames from experiment 1 of the 5ft + mutual gaze + no talk condition (left) and the 10 ft no gaze + talk condition (right). B. Breaching probabilities from experiment 1. Dots represent the estimated marginal mean and lines represent 95% confidence intervals. Empirical breaching rates are represented with white diamonds. C. Video frames from experiment 2 of the 45 degrees offset facing condition (left) and face to face condition (right). D. Breaching rates from experiment 2 plotted by condition and pedestrian walking direction. Arrowheads represent walking pedestrian direction. Conditions are depicted on the right side of the plot. E. Pedestrian trajectories from experiment 2 are plotted and colored by direction. Green is rightward. Black is leftward. Actor body orientation is depicted by red arrows. Types of terrain are represented by the colored background. Offwhite depicts a pedestrian pathway, yellow depicts patches of dirt and vegetation, and grey depicts an asphalt roadway.

Fig. 3 | A. Video frames from experiment 3 of the 10ft + mutual gaze + sign (left) and the 5 ft + no gaze + mural condition (right). B. Breaching probabilities from experiment 1 & 3 at 5 feet (top) and 5 feet (bottom). Dots represent the estimated marginal mean and lines represent 95% confidence intervals. Empirical breaching rates are represented with white diamonds.

Fig. 4 | A. Video frames from experiment 4. B. Pedestrian trajectories from experiment 4 colored by pedestrian gender (blue for man, orange for woman). Colored circles mark where each participant changed course around the actors..

Fig. 5 | On December 7th, 2023 at 12:31:35, over the span of 35.5 seconds, a crowd of 46 pedestrians walked past an art mural while two actors silently stood 10 feet apart while facing and staring at each other. The first 22 pedestrians walk around the actors. 13.6 seconds after the arrival of the first pedestrian, a group of 3 pedestrians (in red) walk between the actors within 0.1 seconds of each other. Just 2.1 seconds later, the first of 12 of the 19 remaining pedestrians also walk between them. A. Video frames before, during, and after the group of 3 pedestrians breaches. B. Pedestrian trajectories during and after the group of 3 pedestrians breaches.

Fig. S5 | Our trajectory inference method with examples from experiment 2. (1) Results of person detection & pose estimation are shown. (2) Results of pedestrian tracking are shown with raw pixel-coordinate trajectories shown in blue. (3) A screenshot of our manual annotation tools. (4) To calibrate our camera, we needed to know its intrinsic properties (how the lens distorts light) and where the camera was located in space. Landmarks used in study 4 are shown in magenta. (5) Raw world-coordinate trajectory data is shown in blue. (6) Smoothed and interpolated data is shown in blue.