Scale-Accurate 3D Vehicle Point Cloud Extraction from Single-Camera Traffic Video

Reliable data extraction is essential to achieving a high-level understanding of the processes ongoing in the traffic environment. The ability to extract 3D structural properties of vehicles enables advanced traffic analysis such as vehicle classification and the detection of traffic rule violations, accidents, and near-collisions. A novel method is presented for extracting 3D structural properties of moving vehicles using a single camera. This method operates by tracking features of the vehicle as it travels through the camera’s field of view. Twoframe structure from motion is then used to extract a 3D point cloud representing the structure of the vehicle. In addition, a robust pose estimation algorithm is given for relating the geometry of the street surface to the position of the camera with minimal user interaction. It is shown that the resulting 3D point cloud can be used to accurately approximate the dimensions of vehicles to within a half foot of their true dimensions.