Unsupervised Detection of Entry and Exit Regions from Vehicle Trajectories for Camera-Agnostic Turning Movement Counts

How can transportation agencies automatically generate turning movement counts (TMCs) from existing surveillance camera networks without the prohibitive cost of manual annotation, camera calibration, or prior knowledge of intersection geometry?

The authors propose a six-step unsupervised pipeline that processes raw vehicle trajectories extracted from video. Unlike traditional trajectory clustering methods that require pairwise similarity comparisons—a computationally expensive process that must be re-run for every new batch of data—this approach identifies persistent spatial region polygons. Once these polygons are established during a calibration phase, future trajectories are classified using simple point-in-polygon containment checks, which is significantly more efficient. The pipeline includes steps for filtering tracking artifacts (such as those caused by vehicle occlusion), estimating the region of interest, selecting informative trajectory endpoints, and clustering these points to define entry and exit zones.

The pipeline was evaluated across 25 surveillance cameras in Bengaluru, India, and 10 sequences from the UA-DETRAC benchmark. Through a systematic statistical analysis of over 19,000 pipeline executions, the authors identified a robust, empirically grounded configuration. Under this configuration, the system achieved a median classification error of approximately 3% across all test sites, including 16 held-out locations. The method demonstrated greater stability across diverse camera views and lower computational costs compared to existing trajectory clustering baselines.

Manual TMC collection is a major bottleneck in urban traffic management, often resulting in data scarcity that prevents agencies from optimizing signal timing or performing safety analyses. By enabling fully automated, camera-agnostic region identification, this pipeline allows cities to leverage existing public safety camera infrastructure for traffic engineering, effectively turning thousands of uncalibrated cameras into a scalable, low-cost traffic monitoring network.