FVHNet: Homography Matrix Estimation for Virtual Camera

Zhiming Hou, Yuanzhouhan Cao

Abstract: Robust lane detection in real-time is one of the foundations for advanced autonomous driving, which can provide substantial amounts of useful information for Autonomous Driving Systems, vehicle self-control, localization, and map construction. Since the in/extrinsic parameters of different vehicles are various which has a significant impact on the results of 3D lanes, we introduce the Virtual Camera that unifies the in/extrinsic parameters of cameras mounted on different vehicles to guarantee the consistency of the spatial relationship among cameras. It can effectively promote the learning procedure due to the unified visual space. Different from using fixed intrinsic parameters and extrinsic parameters to transform front-view images to virtual images, we train a network, termed FVHNet, that estimates the parameters of an ideal perspective transformation, conditioned on the input image. In order to train FVHNet for outputting the transformation matrix that is optimal for perspective transformation, we construct a loss function referred FVHLoss. In summary, we propose a method which can unify the in/extrinsic parameters of different vehicles with learnable homography matrix. We verify our method on the Apollo 3D Lane Synthetic Dataset and achieve competitive results.