Abstract
AVP服务会缓和电动车现有两个缺点: 有限的行驶范围和很长的充电时间.
v-charge用相机和超声波在GPS-denied的区域全自动形式. 这篇paper叙述了下述几方面的优势:
- network communication
- parking space scheduling
- multi-camera calibration
- semantic mapping concepts
- visual localization
- motion planning
这个项目推动了视觉定位, 环境感知和自动泊车到厘米级别的精度.
研发的infrastructure-based camera calibration, semi-supervised semantic mapping concepts极大的减少了维护的成本.
1. Introduction
只用了4个鱼眼相机, 两个stereo相机和超声波雷达.
2. Platform and Sensor Setup
3. Multi-Camera Calibration
developed unsupervised, highly accurate calibration methods for the surround view camera system. the calibration method makes use of natural features in the environment to minimise infrastructure setup costs.
4. Offline Mapping
用SfM的方法离线建图. 每一个3D有额外的descriptors from all images.
5. Perception
用SfM pipeline来全方位.
A. Motion Stereo / Structure from Motion
用plane sweeping.
B. Occupancy Grid Map Fusion
6. Semantic Mapping
sec4 建立了一个metric layer of the map stack.
这里用semantic layer扩展了map stack, 其中有三个特别的部分:
- a road graph
- parking space的位置
- a speed profile
A. The Road Graph
通过metric layer计算的pose组成了lanes.
B. The Parking Labels
C. Speed Map
创建了额一个probabilistic graphical model用路线的位置和parking space作为先验.
7. Communication and Scheduling
8. Visual Localisation
在drop-off位置开始已定位. 定位只用了单目的相机和自然特征.
会用不同时间和日子的地图来augment地图. 要重复这个步骤.