2nd/10_Wiki/Topics/AI_and_ML/Automated_Mapping.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-automated-mapping	Automated Mapping (SLAM / HD Map)	10_Wiki/Topics	verified	self	자동 매핑 SLAM HD map point cloud bundle adjustment loop closure 3D reconstruction NeRF	none	A	0.9	applied	slam hd-map lidar point-cloud bundle-adjustment loop-closure robotics autonomous-vehicles nerf 3d-reconstruction		2026-05-10	pending	language framework C++ / Python ROS / Open3D / COLMAP / OpenVSLAM

Automated Mapping

📌 한 줄 통찰

"매 unknown 의 explore + 매 self-localize 의 simultaneous". 매 SLAM (Simultaneous Localization and Mapping). 매 sensor (LiDAR, camera, IMU) 의 fusion. 매 robotics / AV / AR / VR 의 spatial intelligence 의 base. 매 modern: 매 NeRF / Gaussian Splatting 의 photoreal map.

📖 핵심

매 SLAM 의 4 stage

1. Sensor data: LiDAR / camera / IMU / GPS.
2. Feature extraction: ORB, SIFT, SuperPoint, LoFTR.
3. Pose + map estimation: 매 EKF / particle filter / graph.
4. Loop closure + global optimization: 매 bundle adjustment.

매 SLAM type

Visual SLAM

• 매 camera only.
• 매 ORB-SLAM3 (state-of-the-art classic).
• 매 DROID-SLAM (deep learning).

LiDAR SLAM

• 매 point cloud.
• 매 LOAM, LeGO-LOAM, FAST-LIO.
• 매 sparse + accurate.

Visual-Inertial (VIO)

• 매 camera + IMU.
• 매 VINS-Fusion, OpenVINS.
• 매 robotics, AR/VR.

LiDAR-Visual-Inertial

• 매 multi-sensor fusion.
• 매 LIO-SAM, FAST-LIVO.

매 핵심 component

Front-end

• 매 feature extraction.
• 매 matching (RANSAC).
• 매 motion estimation.

Back-end

• 매 graph optimization.
• 매 g2o, Ceres, GTSAM.
• 매 nonlinear least squares.

Loop closure

• 매 same place revisit 의 detect.
• 매 DBoW2, NetVLAD.
• 매 drift 의 correct.

Mapping

• 매 occupancy grid (2D).
• 매 OctoMap (3D voxel).
• 매 mesh / point cloud.

Bundle Adjustment (BA)

• 매 nonlinear optimization.
• 매 reprojection error 의 minimize.
• 매 camera pose + 3D point 의 동시 추정.
• 매 SLAM 의 backbone.

Modern / deep learning

• DROID-SLAM: 매 differentiable.
• NeRF (Neural Radiance Field): 매 photorealistic 3D.
• Gaussian Splatting (3DGS, 2023): 매 fast NeRF alternative.
• NICE-SLAM: 매 dense neural SLAM.
• Gaussian-SLAM.

HD Map (autonomous driving)

• 매 lane geometry.
• 매 traffic sign / signal.
• 매 routing graph.
• 매 cm-level accuracy.
• 매 update mechanism.

매 응용

1. Autonomous vehicle: HD map.
2. Drone: indoor + outdoor.
3. AR / VR: room understanding (ARKit, ARCore).
4. Robot vacuum: 매 home map.
5. Indoor robot: 매 warehouse, 매 hospital.
6. Surveying: 매 building, 매 mine.
7. Underwater: 매 sonar + visual.
8. Photogrammetry: 매 cultural heritage.

매 challenge

1. Dynamic objects: 매 person, vehicle.
2. Featureless environment: 매 white wall.
3. Lighting: 매 dark / bright extremes.
4. Long-term map: 매 changing environment.
5. Scale ambiguity (monocular): 매 metric scale.
6. Computational cost: 매 real-time.

💻 패턴

ORB-SLAM3 (C++)

# 매 build
mkdir build && cd build && cmake .. && make -j8

# 매 run with EuRoC dataset (visual-inertial)
./Examples/Stereo-Inertial/stereo_inertial_euroc \
    Vocabulary/ORBvoc.txt \
    Examples/Stereo-Inertial/EuRoC.yaml \
    /path/to/V1_01_easy \
    Examples/Stereo-Inertial/EuRoC_TimeStamps/V101.txt

Python visual SLAM (pyslam-style)

import cv2
import numpy as np

class SimpleVO:
    def __init__(self, K):
        self.K = K  # 매 camera intrinsic
        self.orb = cv2.ORB_create(2000)
        self.matcher = cv2.BFMatcher(cv2.NORM_HAMMING)
        self.prev_kp, self.prev_des = None, None
        self.pose = np.eye(4)
    
    def process(self, frame):
        kp, des = self.orb.detectAndCompute(frame, None)
        if self.prev_des is None:
            self.prev_kp, self.prev_des = kp, des
            return self.pose
        
        matches = self.matcher.match(self.prev_des, des)
        matches = sorted(matches, key=lambda x: x.distance)[:200]
        
        pts1 = np.array([self.prev_kp[m.queryIdx].pt for m in matches])
        pts2 = np.array([kp[m.trainIdx].pt for m in matches])
        
        E, mask = cv2.findEssentialMat(pts1, pts2, self.K, cv2.RANSAC, 0.999, 1.0)
        _, R, t, _ = cv2.recoverPose(E, pts1, pts2, self.K, mask=mask)
        
        T = np.eye(4)
        T[:3, :3] = R
        T[:3, 3:] = t
        self.pose = self.pose @ T
        
        self.prev_kp, self.prev_des = kp, des
        return self.pose

Open3D (point cloud)

import open3d as o3d

# 매 load + visualize
pcd = o3d.io.read_point_cloud('scan.ply')
o3d.visualization.draw_geometries([pcd])

# 매 ICP registration
source = o3d.io.read_point_cloud('scan1.ply')
target = o3d.io.read_point_cloud('scan2.ply')

result = o3d.pipelines.registration.registration_icp(
    source, target,
    max_correspondence_distance=0.5,
    estimation_method=o3d.pipelines.registration.TransformationEstimationPointToPoint(),
)
print(result.transformation)

COLMAP (photogrammetry)

# 매 image set → 매 3D reconstruction
colmap automatic_reconstructor \
    --workspace_path /path/to/workspace \
    --image_path /path/to/images

NeRF (instant-NGP)

import tinycudann as tcnn
import torch

# 매 hash grid encoding (instant-NGP)
encoder = tcnn.Encoding(n_input_dims=3, encoding_config={
    'otype': 'HashGrid',
    'n_levels': 16,
    'n_features_per_level': 2,
    'log2_hashmap_size': 19,
    'base_resolution': 16,
    'per_level_scale': 1.5,
})
mlp = tcnn.Network(n_input_dims=encoder.n_output_dims, n_output_dims=4, network_config={
    'otype': 'FullyFusedMLP', 'activation': 'ReLU',
    'output_activation': 'None', 'n_neurons': 64, 'n_hidden_layers': 2,
})

def render(rays_o, rays_d):
    samples = sample_along_rays(rays_o, rays_d)
    encoded = encoder(samples)
    rgb_sigma = mlp(encoded)
    return volume_render(rgb_sigma, samples)

Gaussian Splatting (3DGS, 2023)

# 매 SfM 의 result 의 import
python train.py -s /path/to/colmap-output -m /path/to/output

# 매 view interactive
./SIBR_remoteGaussian_app -m /path/to/output

Loop closure (DBoW3)

#include <DBoW3/DBoW3.h>

DBoW3::Vocabulary vocab("ORBvoc.bin");
DBoW3::Database db(vocab, false, 0);

// 매 keyframe 마다 add
DBoW3::BowVector bow;
vocab.transform(descriptors, bow);
db.add(bow);

// 매 query: 매 매 frame 의 lookup
DBoW3::QueryResults ret;
db.query(bow, ret, 5);
if (ret[0].Score > 0.7) {
    // 매 loop closure detected!
}

🤔 결정 기준

상황	Approach
Indoor robot	Visual-Inertial (ORB-SLAM3)
Outdoor AV	LiDAR + camera + IMU + GPS
AR (mobile)	ARKit / ARCore
Photoreal 3D	Gaussian Splatting
Photogrammetry	COLMAP
Drone outdoor	VIO + GPS
Robot vacuum	LiDAR 2D SLAM
Photoreal AR	NeRF / 3DGS

기본값: Visual SLAM = ORB-SLAM3. LiDAR = LIO-SAM. Photoreal = Gaussian Splatting.

🔗 Graph

• 부모: Robotics · Computer-Vision · Spatial-Computing
• 변형: Visual-SLAM · LiDAR-SLAM · VIO · Visual-Inertial-SLAM
• 응용: Autonomous-Vehicles · AR-VR · Drone · HD-Map · Photogrammetry
• Modern: NeRF · Gaussian-Splatting · DROID-SLAM · NICE-SLAM
• Adjacent: Bundle-Adjustment · Loop-Closure · Bayesian-Brain-Hypothesis

🤖 LLM 활용

언제: 매 robot navigation. 매 AR/VR system. 매 3D reconstruction. 매 AV mapping. 언제 X: 매 2D image processing only. 매 single static image (use SfM).

❌ 안티패턴

• Pure visual outdoor (no IMU): 매 fast motion 의 lose.
• No loop closure: 매 drift 폭발.
• Static map assumption (urban): 매 dynamic obj 의 noise.
• Featureless environment: 매 SLAM fail (LiDAR 의 fall back).
• Offline only: 매 real-time latency 의 ignore.
• No relocalization: 매 lost 시 의 recovery X.

🧪 검증 / 중복

• Verified (ORB-SLAM3, FAST-LIO, NeRF, 3DGS papers).
• 신뢰도 A.
• Related: Autonomous-Vehicles · Computer-Vision · Robotics · NeRF · Gaussian-Splatting.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — SLAM type + ORB-SLAM3 + Open3D + NeRF + 3DGS code