Paper Study 3-3 / 논문 스터디 3-3

Paper Study 3-3

영역 · 자원: Rokey BootCamp (https://www.notion.so/Rokey-BootCamp-1ad39aa2512b803ab013efcd70fb17d9?pvs=21) 고정하기: No 프로젝트: Rokey study 3-3 (https://www.notion.so/Rokey-study-3-3-1bb39aa2512b80148ed0c3594b835956?pvs=21)

KISS-SLAM (2025.03.16)

A Simple, Robust, and Accurate 3D LiDAR SLAM System With Enhanced Generalization Capabilities

• https://arxiv.org/pdf/2503.12660
• https://github.com/PRBonn/kiss-slam

KISS : Keep It Small and Simple

---

• LiDAR-only SLAM not IMU, GNSS … but high-performing
  • pose accuracy
  • map
  • real world possible
• Graph-based SLAM & LiDAR odometry
• match local maps → optimize the trajectory in a pose graph optimization step

Our Approach to LiDAR-based SLAM

---

LiDAR Odometry Using KISS_ICP

---

https://arxiv.org/pdf/2209.15397

• The LiDAR sensor in the odometry frame at time $t$, : $T_t\in SE(3)$
• The incoming point cloud : $P=\{p_i|p_i\in {\mathbb{R}}^3\}$
• The first pre-process by de-skewing and voxel downsampling : $S=\{s_i|s_i\in {\mathbb{R}}^3\}$

(https://velog.io/@se0yeon00/PointCloud-를-Voxel-grid-로-다운샘플링-하기)

• The previous estimate of the LiDAR pose : $T_{t-1}$
• A constant velocity motion model prediction : $△T_t\in SE(3)$
• An initial guess for the current LiDAR pose :

$${\hat{T}}_t=T_{t-1}△T_t$$

→ ICP algorithm : point-to-point Iterative Closet Point

(https://alida.tistory.com/105)(https://nbviewer.org/github/niosus/notebooks/blob/master/icp.ipynb#Non-linear-Least-squares-based-ICP)

• The local map points : $\phi =\{q_i|q_i\in {\mathbb{R}}^3\}$
• residual : $r$,

$$r(T)=Ts-q$$

• cost function : ( c : The set of nearest neighbor correspondenecs)

$$\chi (\widehat{T_t}=\sum _{(s,q)\in c}{\Vert r(\widehat{T_t})\Vert }_2^2$$

• minimize EQ (least squares fashion) :

$$△w=\underset{△w}{\mathrm{argmin}}\chi (\widehat{T_t}⨁△w),△w\in {\mathbb{R}}^6$$

Local Mapping and Pose Graph Construction

---

→ approach avoids main training a single global map through a local map-splitting strategy.

• Local map : $M_k$
• Keypose : $T_k\in SE(3)$
• local trajectory : ${\tau }_k\in \{T_{t->t+1}|t_s\le t<t_e\}$
• voxel grid : ${\nu }_k$

→ Create a new local map when the traveled distance exceeds a certain threshold $\beta$

Loop Closing

---

https://arxiv.org/pdf/2501.07399

• Ground Alignment → x-y-plane of the keypose reference frame
• BEV Density Image
• ORB Feature Descriptors
• the SzymkiewiczSimson overlap coefficient: → 2D alignment of the density images
  • compute mean of the point coordinates $u_i$ and a per-voxel normal vector $n_i$
  • → $N_k=\{\{u_i,n_i\}|u_i,n_i\in {\mathbb{R}}^3\}$ (resulting point cloud)
  • apply the initial guess ${\hat{T}}_{i->j}$ & perfom a point cloud registration step between the voxel-based point clouds $N_i,N_j$

$$\Gamma (N_i,N_j,T_{i->j})=\frac{|N_j\cap T_{i->j}\oplus N_i|}{min(|N_j|,|N_i|)}$$