The Walking Behaviour of Pedestrian Social Groups and Its Impact on Crowd Dynamics

Here’s an overlooked fact about crowds that changes how you see a busy street: most people aren’t walking alone. They’re walking with someone. Moussaïd and colleagues watched real sidewalks and found that in many public spaces, the dominant unit isn’t the individual but the group. Two to four people, moving together, talking, glancing at each other. And once you notice that, a big question arises: how does the need to stay connected inside a group reshape the way the whole crowd flows? They set up cameras in two natural settings a year apart—one in Toulouse on a workday afternoon, and one on a crowded shopping walkway on a Saturday. From hours of public video, they manually tracked positions and identified who was walking with whom, using social cues like talking, laughter, and gestures. In the quieter site, density was low, around three hundredths of a person per square meter. On the busy walkway, it was higher, a quarter per square meter. Across both, the dataset was big—thousands of people, more than a thousand groups—and the basic picture was unmistakable: groups dominate. About 55 percent of pedestrians in the low-density site were in groups, and that jumped to roughly 70 percent in the denser, leisure setting. If you’ve ever weaved through a shopping street and felt the sidewalk get wide and then suddenly narrow, that’s partly this—groups spreading out when they can, compacting when they must. Even the way group sizes are distributed tells a story. In the lower-density site, groups of one, two, three, and four people followed a classic zero-truncated Poisson pattern with a rate around 0.83. That’s a statistical way of saying "mostly pairs and trios, with fewer larger clumps," and it fit reasonably well. In the busier shopping scene, there was a twist: fewer solo walkers than a simple Poisson would predict and more pairs than expected, a significant shift. Context matters. On a Saturday afternoon, you don’t browse alone as often as you commute alone. Speed is where the social rubber meets the road. Two forces pull in opposite directions. If you’re an isolated pedestrian, you optimize for getting there. If you’re in a group, you also optimize for staying in touch. The data show both effects at once: higher overall speeds at low density and a steady, independent slowdown as groups get larger. Quantitatively, each extra person in a group trims the pace. In the low-density site, think of starting near 1.26 meters per second and losing about 0.04 meters per second with each additional member. In the busier setting, you start around 1.24 meters per second and lose about 0.08 meters per second per person. Statistically, those slopes aren’t distinguishable across densities by analysis of covariance; they point in the same direction. More people in your cluster means a slower walk, regardless of how crowded the street is. Now, that slowdown isn’t random. It comes with a beautiful, density-driven choreography you can see if you watch groups from above. At low density, two friends tend to go shoulder to shoulder, almost in a straight line across the sidewalk, at roughly right angles to their direction of travel. Their spacing is comfortable—about 0.78 meters apart in the quieter site. Three or four people, same story: they fan out roughly horizontally when there’s space. As the environment fills, you see something else. Three-person groups bend into a forward-slanted V: the middle person slightly back, the two flanking members angled inward. In the data from the denser site, the angle on one side opens up to a bit over a hundred degrees, while the other side compresses toward seventy, a telltale asymmetry that points the group like an arrow. Distances tighten, too, to around half a meter. Four-person groups do a similar trick, curving into a shallow U so that everyone keeps everyone else visible. That’s the thread running through all of it: mutual visibility and an easy line of conversation. To see how far that thread could be pulled, Moussaïd and colleagues took the classic social force model. It treats walking as the result of a desired velocity plus repulsions from walls and other people, and they added a new term that encodes the social glue inside a group. The baseline mechanics stay familiar. Each pedestrian relaxes toward a desired speed and direction over a characteristic time. If you want numbers, think of a target speed around 1.3 meters per second and a relaxation time of about half a second. Walls push you away with a force that falls off exponentially with distance; in their setup, that repulsion extends over a few tenths of a meter and is strong up close. Add in the usual person-to-person repulsions to keep bodies from overlapping. The new ingredient is gaze. Not eye movements, literally, but the idea that you carry a cone of attention, a field of view extending about ninety degrees left and right from your current gaze direction, and you prefer to keep the rest of your group inside it. In the model, each person constantly estimates where the center of mass of the other group members sits relative to them. If it drifts outside their visual field, they would need to rotate their head by some angle to see it. Big rotations are uncomfortable if you’re trying to keep walking, so the model translates that into a kind of braking force: the larger the implied head turn angle, the stronger the deceleration. A single parameter, b1, dials how much you care about keeping your partners in view while you walk and talk. Two other pieces round out this group term. There’s a gentle attraction back toward the center of the group, with strength b2. It kicks in only if you’ve strayed beyond a threshold tied to group size—imagine a meter or so scaled by the number of members. And there’s a short-range repulsion, with strength b3. It activates if two members overlap given a safety distance: a hard don’t step on each other rule set by a minimum spacing of about 0.8 meters. Put all three together—the gaze-driven deceleration, the center-of-mass pull, and the internal repulsion—and you get a force that keeps groups cohesive and mutually visible without letting them collapse into a single blob. Here’s the punchline from the simulations. Turn off the gaze term by setting b1 to zero and groups behave like loose flocks that merely try to stay together. They fall into an inverse V pattern that’s aerodynamically tidy for the flow and moves at nearly the same speed as isolated walkers. Switch b1 to the empirically calibrated value—b1 equals four—and everything snaps into the forward V and shallow U shapes you saw on the street. There’s a cost. Relative to isolated pedestrians, the modeled crowd’s average speed drops by about seventeen percent. In the model’s speed-density curves, that gap is small when the environment is empty—fast progression and social coordination aren’t fighting each other yet—but widens at moderate densities as the need to keep everyone in view reshapes each group. The match between model and reality runs deeper than a visual rhyme. They calibrated those social parameters by scanning the space of b1, b2, and b3 to best fit the observed geometry inside groups—those average angles and distances I mentioned. They then checked, with t-tests, that the simulated and observed angle distributions were basically indistinguishable. The best-fit numbers paint a coherent picture: b1 equals four, b2 equals three, b3 equals one, with a ninety-degree field of view and the minimum spacing set to 0.8 meters. They dropped these groups into two simulated street grids sized to mirror the real sites. One was about eighteen by eighteen meters for the sparse setting, while the other was five by fourteen for the crowded one. They randomized starting positions and directions, and ran the whole thing a thousand times per condition. After ten seconds of warm-up and five seconds of measurement, the averages lined up. The forward-leaning V and U patterns emerged at higher density. Side-by-side lines appeared at low density. And they observed the same linear drop in speed with each extra person in a group. Empirically, the group-size slowdown didn’t depend on density in a statistically meaningful way, and the model reproduced that as well. If you like thinking in variables, you can summarize the movement of a person in this framework as a tug-of-war among four forces. One points you toward your goal at your preferred pace, pulling you back when you drift from that target. One pushes you away from walls, getting weaker as you move farther from the boundary. One is the usual repulsion from other bodies at close range. And one is the social force that keeps your group in sight and at a comfortable distance—penalizing large head turns with a speed cost, nudging you inward if you drift out, and keeping you from overlapping. The interesting bit is where that last one wins. At low density, the first and fourth forces are aligned. At moderate density, the social term starts to bend trajectories into those convex V and U shapes, and, when you add it up over hundreds of groups, you get a measurable penalty in throughput. There are a few other nuances worth mentioning before we zoom out. The spatial metrics show that at low density, two-person groups almost lock to a right angle with about three-quarters of a meter between them. In the crowded site, that separation compresses to roughly half a meter. In three-person groups at higher density, one side of the triangle opens wide while the other narrows—a geometry that pulls the whole trio into a forward slant so all faces stay readable. And looking beyond the calibrated cases, the researchers note where this framework might bend. At very high densities, you start to see single-file, river-like formations as physical constraints overwhelm social preferences. Very large groups may split naturally. And who is talking can bias who walks where, a pattern consistent with earlier sociological work on conversations in motion. Why does this matter? Because it means capacity isn’t just a number you compute from individual stepping speeds. It’s shaped by how we insist on staying together. Urban planners, event organizers, transit designers—they’re all modeling flows. If most of those flows are actually groups with a built-in tendency to spread, bend, and slow for the sake of communication, designs that ignore that social force will overestimate how fast a crowd can move and how much space it needs. The flip side is optimistic: plan for groups, and you can create more realistic, safer environments without blaming people for acting like people. If you want one takeaway to carry with you on your next walk, make it this: the V-shaped trio in front of you isn’t being inefficient. They’re doing a tiny dance of coordination that lets them keep eye contact and a sentence going while the world slides by. Moussaïd and colleagues showed that dance is systematic, predictable, and it changes the whole river of motion around it—by about seventeen percent in speed when social interaction is strong. That’s the price of conversation, and it’s a price crowds gladly pay. This lecture was created by ennepō. Go to ennepo dot A I to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.