Browsing by Author "Li, Yingmao"
Now showing 1 - 1 of 1
- Results Per Page
- Sort Options
Item Multiple Lane Boundary Extraction, Model Fitting and Tracking in Presence of Heavy Outliers, and Predictive Reconstruction of 3D Lane/Road Geometry from a Single Image(2018-12) Li, Yingmao; Gans, NicholasThere are a variety of applications that can benefit from the knowledge of accurate 3D road geometry, which can provide information for driver assistant systems on autonomous vehicles as well as unmanned aerial systems. Existing solutions are mostly based on either range sensors such as a LiDAR, or triangulating through multiple image frames, such as a stereo camera, or structure from motion. However, knowing the 3D road geometry from one single image posts many advantages. For example, it allows the mobile platforms to work without being constrained by payload, energy consumption, cost, etc. Furthermore, it provides the state of the mobile platform directly without accumulating error over time, which is an appealing feature for applications such as simultaneously localization and mapping (SLAM). We present an algorithm to reconstruct the 3D road geometry accurately from one single frame of image valid for all directions of the road curves. The testing results over a public dataset shows that our algorithm is able to achieve a comparable accuracy against a high-end LiDAR sensor over a long distance, given the knowledge of the road width. There are two major contributions in our work. Our first contribution is a highly efficient model fitting and tracking framework that combines the random sample consensus (RANSAC) algorithm and (Extended) Kalman filter together to work in presence of heavy noise and outliers by leverage the advantages of both algorithms. The second contribution is that we propose a single view, single image frame 3D road geometry reconstruction algorithm that is formulated as an optimization problem with the 3D road curvature model as the constraint. We integrate them together as an effective system, and show the feasibility of our algorithms using both simulation and real world experiments.