Skip to main content
Back

AIR-VC Counts Traffic 4x Faster by Finding Its own Camera Sweet Spot

A new framework automatically picks the best counting line in traffic video, hitting near-perfect accuracy while running up to four times faster than full-frame processing.

  • AIR-VC automatically selects the optimal counting line and ROI for vehicle counting from fixed traffic cameras.
  • Combines three signals: detection confidence, tracking confidence, and detection density heatmaps to avoid blind spots.
  • Achieves up to 403% speed improvement over full-frame processing while hitting 100% accuracy on several benchmarks.
  • Evaluated on UA-DETRAC, GRAM, CDnet 2014, and ATON across rain, night, and heavy occlusion.
  • Detector-agnostic design tested with YOLOv8 and DeepSORT for detection and tracking.
  • No public code released; framework is line-based and does not yet handle intersections or turning counts.

Counting cars from a fixed traffic camera sounds trivial until you actually try it. Trackers swap identities between vehicles, occlusions eat detections, and the region of the frame that gives the cleanest count differs for every camera angle. A new paper from researchers at Aswan University and Universidad Francisco de Vitoria proposes AIR-VC, an Automated, Intelligent, ROI-based Vehicle Counting framework that figures out the best counting zone on its own and then only processes that slice of pixels.

Most videos reach 100% accuracy, processing runs up to four times faster than full-frame inference, and the method holds up on complex multi-road scenes where prior work tended to break.

Why manual counting lines keep failing

Deep-learning vehicle counters typically run detection and tracking on every pixel of every frame, which is expensive. To cut cost, most systems restrict work to a region of interest (ROI) or a virtual counting line drawn across the road. That line is usually placed by hand, or hard-coded to a heuristic like the frame's center, the area closest to the camera, or a predefined ROI.

The authors show why this fails in practice. Accuracy curves obtained by sliding counting lines across three videos vary significantly between scenes, meaning no universal optimal ROI exists. Within a single video, though, different sequences show similar patterns, so a fixed scene does have a stable sweet spot. Prior automated attempts based on highest detection confidence, motion masks, or texture complexity all generalized poorly because the best counting region is not always the most confident or the most congested.