Grid-Based LiDAR SLAM for Dynamic Environments

This repository contains the implementation and evaluation of a grid-based LiDAR SLAM approach with dynamic object filtering in a TurtleBot3 simulation environment. The experiments focus on mapping and localization performance under dynamic conditions.

---

📦 Prerequisites

• Ubuntu 22.04
• ROS 2 Humble installed and sourced
• turtlebot3_description, turtlebot3_gazebo, and other core TurtleBot3 packages

---

🔧 Installing Dependencies

1. Update apt cache

   sudo apt update

2. Install extra system deps
   Make sure you have apt_requirements.txt in the workspace root containing:

   xterm
   

   Then run:

   xargs sudo apt install -y < apt_requirements.txt

3. Install all ROS 2/Gazebo/TurtleBot3 deps

   sudo rosdep init || true
   rosdep update
   rosdep install --from-paths src --ignore-src -r -y

---

Materials

Experimental Environment and Robot Platform

Simulation Environment: turtlebot3_world

All experiments were conducted in the pre-built TurtleBot3 simulation environment (turtlebot3_world) provided by the TurtleBot3 ROS packages. A moving cubed object was added with customizable pose, velocity, and limits.

TurtleBot3 Models Comparison

TurtleBot3 comes in three variants: Burger, Waffle, and Waffle Pi. Key specifications relevant to SLAM are summarized below:

Specification	Burger	Waffle	Waffle Pi
Processor	OpenCR + SBC	OpenCR + SBC	OpenCR + Raspberry Pi 3
LiDAR Sensor	LDS-02 (360°)	LDS-02 (360°)	LDS-02 (360°)
Camera	None	Intel RealSense	Raspberry Pi Camera
IMU	Yes	Yes	Yes
Wheel Encoder	Yes	Yes	Yes
Size (W×L×H, mm)	138×178×192	281×306×141	281×306×141
Weight	~1.0 kg	~1.8 kg	~1.8 kg
Battery Life	~2.5 hours	~1.5 hours	~1.5 hours

Note: Our experiments focus solely on LiDAR-based SLAM and grid-based dynamic object filtering. We used the Burger model, which has no camera, to reduce computational overhead.

---

Methods

SLAM Toolbox Custom Integration

We customized the SLAM Toolbox launch files and YAML configurations for ROS 2 Humble + Gazebo Classic.

Launch File Customizations

• Centralized Config Loading: Both launch/online_async_launch.py and launch/online_sync_launch.py accept a slam_params_file argument pointing to our custom YAML configuration.
• Async vs Sync Choice: Separate launch files allow switching between asynchronous and synchronous mapping modes easily.

YAML Configuration Modifications

• Custom Scan Topic: Changed scan_topic from /scan to /preprocessed_scan to include preprocessing.
• LiDAR Range Settings: Set min_laser_range and max_laser_range to sensor specifications.
• Map Update Interval: Tuned for CPU efficiency (1.0 s async, 0.5 s sync).
• Zero Travel Thresholds: Set minimum_travel_distance and minimum_travel_heading to 0.0.

Median Filter

A sliding-window median filter is applied to remove spurious dynamic points:

$$ y_i = \mathrm{median}\Big(x_{i - \lfloor w/2 \rfloor}, \dots, x_i, \dots, x_{i + \lfloor w/2 \rfloor}\Big) $$

Observations:

• Reduces transient dynamic artifacts.
• Cannot remove slow-moving or large objects perfectly.
• Improves map quality even with small window sizes.

Metric	Before Filtering	After Filtering
Dynamic SSIM	0.8668	0.9032
Normalized NE (px)	0.35	0.11

Consecutive-Scan Background Subtraction

• Projects scans to world coordinates and compares consecutive scans.
• Limitation: Fails due to one-frame memory, no proper registration, and high computational cost.

Map Comparison Filtering

• Detects dynamic cells by differencing consecutive occupancy grids.
• Limitation: SLAM Toolbox does not update free cells immediately, leading to missed detections.

Occupancy Threshold Manipulation

• Adjusted occupancy threshold ($\tau$) to suppress dynamic artifacts.
• Optimal value: ($\tau = 0.4$) balances dynamic suppression with static structure retention.

Engineering Experiment

We evaluated median filtering and occupancy thresholding across dynamic scenarios:

Group	Experiments	Description
Baseline	4	No filtering
Median	4	Sliding-window median filtering
Occupancy	4	Thresholding occupancy probabilities
Static	-	Ground-truth reference map

Experiment #	x [m]	y [m]	z [m]	v [m/s]
1	0.4	0.6	0.2	0.05
2	0.4	0.6	0.2	0.5
3	0.4	0.2	0.2	0.5
4	0.4	0.2	0.2	0.05

Data Collected:

• Ground-truth robot pose from Gazebo.
• Estimated robot pose from SLAM system.
• Generated maps and trajectories for evaluation.

Map Evaluation Metrics:

• Pixel-wise differences
• Free/occupied classification: Precision, Recall, F1-score

---

Results and Discussion

Map Evaluation

Pixel-Level Differences:

• Median filter experiments: ~1.5%–2.1% difference.
• Occupancy threshold filter: up to 2.6%, removing most dynamic objects but causing some static deletion.

F1 Scores:

• Median filter: up to 0.9933, best balance.
• Occupancy filter: slightly lower F1, high recall but lower precision.

Discussion:

• Occupancy threshold removes dynamic objects effectively but misclassifies some static areas.
• Median filtering preserves structure but may retain minor motion artifacts.
• Choice depends on application-specific trade-offs.

Localization Accuracy Analysis

• Median filter experiments (m5–m8) reduce translational and orientation errors compared to static baseline.
• Occupancy filter shows mixed localization improvement across experiments.

Discussion:

• Both filtering methods can improve localization depending on scenario.
• Trade-offs between dynamic object suppression and static map accuracy must be considered.

---

References

• ROS 2 SLAM Toolbox: https://github.com/SteveMacenski/slam_toolbox
• TurtleBot3: https://emanual.robotis.com/docs/en/platform/turtlebot3/overview/