Preventing Static Element Double Counting in Urban Video Segmentation for Autonomous Driving

Autonomous driving stacks often summarize long drives into scene-level metrics like how much road, sidewalk, or building surface was seen. When a route loops or the car revisits a view, static surfaces get counted again, which inflates composition figures and skews downstream analytics. This paper defines that failure mode and tackles it directly. We propose a revisitation-aware framework that blends three parts: temporally smoothed per-pixel labels, pixel-level instance tracking, and a semantics-driven revisitation detector with spatial verification to recognize previously seen static regions and suppress duplicate tallies. The design is simple and modular, so it drops into standard backbones without fuss. On looped sequences from Cityscapes and KITTI, the method improves route-level composition accuracy, cutting the Static Overcount Ratio from about 2.10 to 1.05 while maintaining or improving frame mIoU and multi-object tracking quality. We report composition with the Static Overcount Ratio and Unique Static Coverage, alongside standard segmentation and tracking metrics, so it is clear when revisits are handled correctly. Why it matters. Counting static elements only once per physical surface makes route summaries more truthful. That helps autonomous-driving analytics, urban mapping, and planning pipelines that depend on reliable scene composition.