Video Processing Method for High-Definition Maps Generation

This paper proposes a three-phase method for generating high-definition (HD) maps based on the analysis of video recordings obtained from DVRs or smartphones mounted in a vehicle. The method uses a convolutional neural network (CNN) as a classifier of map object types on the video frames. The calculation of the distances while HD-map generating is carried out using a mathematical model taking into account three-dimensional space and tilt angles of the camera. CNN had trained on the KITTI image set to determine the asphalt surface, then it applied to a video record with 1800 frames to generate a layer of HD-map. The test results showed a measurement error is about 5% at a distance of 15 meters from the camera, thus, the proposed method could be used as a tool for HD-maps generation.