Ground Plane Projection for Improved Traffic Analytics at Intersections

TL;DR

The paper tackles accurate turning movement counting at intersections by arguing for trajectory analysis on the ground plane rather than the image plane. It introduces a three-phase pipeline (calibration, unsupervised learning, and inference) and deploys three datasets, including new low-vantage and multi-view collections, with ground-truth TMCs and ground-plane mappings. The approach uses unsupervised exemplar-based and KDE-based probabilistic models to classify turning movements, showing that ground-plane classification generally outperforms image-plane methods, and that simple weak fusion across multiple cameras yields further gains. The work demonstrates practical benefits for ITS applications, highlighting improved accuracy in single-view scenarios and meaningful gains from incorporating additional views, with directions for further improvements such as 3D detection and terrain-aware back-projection.

Abstract

Accurate turning movement counts at intersections are important for signal control, traffic management and urban planning. Computer vision systems for automatic turning movement counts typically rely on visual analysis in the image plane of an infrastructure camera. Here we explore potential advantages of back-projecting vehicles detected in one or more infrastructure cameras to the ground plane for analysis in real-world 3D coordinates. For single-camera systems we find that back-projection yields more accurate trajectory classification and turning movement counts. We further show that even higher accuracy can be achieved through weak fusion of back-projected detections from multiple cameras. These results suggeest that traffic should be analyzed on the ground plane, not the image plane

Figures (7)

• Figure 1: Examples from our three datasets. Left: Camera views. Right: Orthophotos from Google Maps. The Region of York dataset (c) provides camera views from each of the four corners of the intersection; only one is shown here.
• Figure 2: Three steps of geometric calibration: (a) Intrinsic calibration; (b) Extrinsic calibration; (c) Region of interest annotation.
• Figure 3: Unsupervised learning pipelines for (a) camera plane and (b) ground plane systems.
• Figure 4: Example representative prototype tracks selected using our unsupervised clustering method, for two CityFlow V2 sites. One track is selected for each lane within each turning movement class. Each class is represented by a different colour. Top: image plane. Bottom: orthophoto.
• Figure 5: Average log likelihood of training tracks for the second half of the training partition for one Trans-Plan site based on KDE model derived from the first half of the training partition. The likelihood peaks at 9.7 pixels in the image plane (left) and 3.3 metres in the ground plane (right).