Blog · arXiv Analysis · Last reviewed July 10, 2026

The Street Video Becomes the Informality Meter

CommuniWave turns street video into a Degree of Informal Behavior score. The governance question is whether a planning office can measure spontaneous public life without turning it into a quiet enforcement target.

Published: July 10, 2026 Modified: July 10, 2026 Last reviewed: July 10, 2026 Tags: arXiv, urban surveillance, computer vision, informal behavior, AI governance

The Paper

The paper is CommuniWave:A Machine Learning Model for Quantifying the Degree of Temporary Informal Behavior in Urban Communities, arXiv:2607.08554 [cs.AI]. The arXiv API lists Hongye Yang, Shien Liu, and Zhihao Xie as authors, with version 1 submitted on July 9, 2026. The arXiv record describes a 17-page, 4-figure paper presented at ASCAAD 2024.

This belongs near AI governance, AI audit trails, surveillance capitalism, urban intelligence, and surveillance critique. The fresh angle is not another dashboard for city managers. It is the moment a community's informal life becomes a computer-vision variable.

What It Measures

CommuniWave is built to detect and score what the authors call the Degree of Informal Behavior, or DIB, in urban communities. The system has two major parts. Behavior Capture Net, or BCN, combines mmaction2 with a YOLOv10-based model named YLX to recognize actions and informal behaviors in street videos. Behavior Eval Model, or BEM, uses random forest regression, with PCA preprocessing, to convert detected features and street-level labels into a DIB score.

The output is a fluctuation chart for streets over time. In the authors' practical test, 10-minute videos from morning, noon, and evening were segmented into 10-second clips, producing 60 predicted DIB values per video. First-class roads and second-class roads showed higher morning DIB values when street vending and pedestrian gatherings appeared; third-class roads had occasional vending and lower overall values in the paper's account.

The Training Frame

The BEM training set used 180 manually annotated 10-second street-video segments from a medium-sized city in southern China. The paper divides the material into three road classes: streets allowing motor and non-motor vehicles, streets not allowing motor vehicles, and pedestrian-only streets, with 60 segments in each class. Cameras were placed 5-10 meters above ground to reduce occlusion and crowd overlap.

Ten urban-community residents, 6 male and 4 female, with an average age of 31.0 years and a standard deviation of 13.89 years, rated the 180 clips on a 1-to-5 temporary-activity scale. The authors say the volunteers signed informed-consent forms for anonymous use of rating data. The mean score, rounded to two decimals, became the label for each clip, and Median Absolute Deviation filtering was used to remove extreme outlier ratings.

YLX was trained to recognize six common informal behaviors in Chinese communities: square dancing, gathering, street vending with equipment, street vending without equipment, three-wheeled motorcycling, and chess playing. The paper says the image data for those labels came mainly from Chinese social media platforms and that the authors curated a 30,000-image YLX dataset with automated annotation followed by manual verification.

Evidence

The paper reports YLX training accuracy improving to 0.794. It also reports BEM Mean Squared Error of 0.9599, Root Mean Squared Error of 0.9798, and an R2 of -0.1681. That negative R2 matters. It means the regression evidence is not a clean story of strong predictive fit on the held-out split, so the paper's stronger planning claims should be read as proposed workflow value, not as mature operational validation.

For an additional unseen-data check, the authors re-shot 10 videos, each 10 seconds long, ran CommuniWave on them, and recruited 10 volunteers to rate the same videos. The Mean Absolute Deviation between model predictions and volunteer average scores was 0.709. The paper describes this as showing some deviation but overall closeness to human ratings.

The SHAP feature-importance analysis is socially revealing. Street vending, with and without equipment, had strong influence across several streets. Three-wheeled motorcycling and walking on the road affected first-class roads. Pedestrian gathering and square dancing mattered on second-class roads. Third-class roads were dominated by street vending in the reported feature account. In other words, the model's signal is not merely a technical category; it is a taxonomy of tolerated, managed, and potentially disputed public presence.

Limits

The authors state several limits directly. BCN cannot cover all categories of informal behavior, and the label library does not generalize automatically across regions, street environments, and sociocultural contexts. The system recognizes surface visual features, so it may miss deeper semantics: the paper gives the example of a model failing to distinguish a car parked on a sidewalk from a car parked in a designated parking space. The authors also note that mmaction and YOLOv10 may crash on some video formats and that researchers may need to manually correct BCN outputs.

The privacy claim also needs a governance receipt, not just a method name. The paper says public-space videos were de-identified to separate personal identity from behavior information. That reduces some risk, but group behavior, location, time of day, vendor activity, and street class can still affect residents and workers when the metric is used for management. A DIB score can become a soft police report if its downstream use is not constrained.

The Receipt

An urban informality receipt should record the camera location and height, filming time, road class, consent basis, de-identification method, retention period, behavior labels, social-media source policy for training images, annotation tool, manual-verification rule, volunteer demographics, rating instructions, outlier filtering, model versions, PCA settings, random-forest parameters, train-test split, YLX accuracy, BEM MSE, RMSE, R2, unseen-video test procedure, Mean Absolute Deviation, SHAP feature report, crash handling, human review process, appeal path for affected groups, and the rule separating observation from enforcement.

The audit question is not whether a city can turn video into a chart. It is whether residents, vendors, dancers, pedestrians, and planners can see how the chart was made, what it misses, who can act on it, and when the metric must stay advisory.

Sources

Return to Blog