Kuan-Chuan Peng
Research Interests
- Computer Vision
- Machine Learning
- 3D Reconstruction
Current Research
Image-based Localization
In certain scenarios common GPS can be not accurate enough, we try to solve this problem using an captured image. This is not a completely new topic and many approaches have been proposed. However, existing approaches can be too complicated and computationally expensive. We address this problem differently: we try to interpret the image in novel and efficient ways, and combine this with noisy GPS reading and map information. In this way, we created a lightweight, but intelligent method for image-based localization.
Mapping for Autonomous Vehicles
To enable a car driving by itself, the first problem that should be solved is accurately localizing it. We try to create a high resolution ground plan texture map that contributes to better vehicle localization, with off-the-shelf common resources: common GPS, ESP encoder, camera, and low resolution satellite image. We propose to identify "anchor points" that associate camera input and the satellite image. Using our method, we are able to localize the vehicle and create high quality ground plan texture map simultaneously.
Indoor Visual Navigation
Project Tango, Semi-dense VO, RGB-D SLAM... we are seeing more and more powerful tools for real-time 3D reconstruction. What is the next step for visual indoor navigation? We are trying to create a system that uses only a camera and an architectural floorplan with no other prior knowledge. Therefore, our method not only solves the navigation problem, but also is practical and inexpensive to be applied in large scale.