Follow the conventional pipeline of SLAM, here are some paper utilizing semantic info.
This kind of methods predict (rather than extract directly) extra information from sensor data.
Given info, the next problem is to determine the correspondence. Robust and efficient data association is the core issue in SLAM.
This type of method focus on providing high-level map for advanced robotics applications.
| Type | Paper | Detial | |
|---|---|---|---|
| Prediciton | Depth | CNN-SVO | 单双目深度估计 |
| Scale | Recovering stable scale in monocular SLAM using object-supplemented bundle adjustment | 尺度不确定性恢复 | |
| Occlusion | EmptyCities | 对被遮挡的场景恢复后再做SLAM | |
| MBO | 去模糊网络 | ||
| Data Association | Learned Feature | CNN-SLAM | 用CNN网络进行特征匹配 |
| AVP-SLAM | 识别环境中的车道线等特征 | ||
| CarFusion/OrcVIO | 用到的结构化特征点LIFT/SuperPoint | ||
| LodoNet | 激光投影到2D平面提特侦匹配 | ||
| Filter Noisy Info | DynaSLAM | 用MaskRCNN来做动态物体删除 | |
| Suma++ | 动态滤除 | ||
| SalientDSO | 显著性应用 | ||
| Lego_LOAM | 滤除植物点 | ||
| Add Constrain | PSF-LO | 几何配准加语义约束 | |
| Probabilistic Data Association for Semantic SLAM | 语义为隐含优化变量 | ||
| Semantic-ICP | 语义为隐含优化变量 | ||
| Integrating Deep Semantic Segmentation Into 3-D Point Cloud Registration | 语义filter | ||
| SIVO | 基于SVO，cubeslam物体级别 | ||
| Deep Matching | SuperGlue | DL解决匹配问题 | |
| Loop Closure/Long Term | X-View | 长期回环检测 | |
| PointNetVLAD | 激光回环 | ||
| Semantically Assisted Loop Closure in SLAM Using NDT Histograms | 含语义的直方图回环 | ||
| GOSMatch | 图匹配的回环 | ||
| Solving | BA-Net | DL方式做优化 | |
| Mapping | With Tag | Kimera | 针对导航任务的SLAM算法结合，主要操作在Mesh上了 |
| Object-level | |||
| Metric | |||
| End-to-End Odometry | DeepVO | 端到端的VO/SLAM |