ORB-SLAM3 Accelerated on NVIDIA Jetson Orin Nano 🚀

An optimized implementation of the ORB-SLAM3 visual SLAM system, designed for embedded edge computing platforms.

This project addresses the high computational latency of visual feature extraction on mobile processors by implementing a Hybrid CPU-GPU Architecture.

---

⚡ Key Features & Optimizations

1. CUDA Accelerated ORB Extraction (GPU)

• Replaced the sequential descriptor extraction with a massive parallel CUDA Kernel.
• Utilizes Constant Memory for ORB patterns to minimize global memory latency.
• Achieves ~5x speedup in the descriptor calculation stage (from 23ms to 4ms).

2. Grid-Based Feature Distribution (CPU)

• Replaced the recursive QuadTree algorithm (standard in ORB-SLAM3) with a linear Grid-based filtering approach.
• Ensures uniform feature distribution with O(N) complexity.
• Reduces CPU overhead and branch mispredictions.

3. Optimized Memory Management

• Implemented Static Memory Pooling on the GPU to avoid cudaMalloc overhead per frame.
• Zero-copy data transfer where applicable (Unified Memory architecture).

---

📊 Performance Results

Tested on NVIDIA Jetson Orin Nano using the KITTI Odometry Benchmark (Sequence 00).

Processing Stage	Original (CPU)	Accelerated (CPU+GPU)	Speedup
Pyramid Building	1.52 ms	2.45 ms	-
Feature Distribution	20.85 ms	11.20 ms	1.86x
Gaussian Blur	2.65 ms	4.10 ms	-
ORB Descriptor	21.40 ms	4.35 ms	4.92x
Data Transfer	2.45 ms	0.45 ms	5.44x
TOTAL FRAME TIME	~49 ms	~22 ms	~2.2x
FPS	20 FPS	45 FPS	Real-Time

Note: While the GPU kernel is 5x faster, the total system speedup is governed by Amdahl's Law, resulting in a 2.2x overall improvement.

---

🛠️ Hardware & Software Requirements

Hardware

• Device: NVIDIA Jetson Orin Nano / Xavier NX / AGX Orin.
• Storage: NVMe SSD (Recommended for high-bandwidth dataset reading).

Software

• OS: Ubuntu 20.04 (JetPack 5.x) or Ubuntu 22.04/24.04.
• Compilers: GCC 9+, NVCC 11.4+.
• Libraries:
  • OpenCV 4.4+ (Must be compiled with CUDA & Eigen support).
  • Eigen3.
  • Pangolin (for visualization).

---

📦 Installation

1. Clone the repository:

   git clone https://github.com/tothantonio/ORB_SLAM3_Accelerated.git
   cd ORB_SLAM3_Accelerated

2. Build the project: We provide a script to handle the mixed C++/CUDA compilation.

   chmod +x build.sh
   ./build.sh

3. Download KITTI Dataset: Download the grayscale odometry sequences from the KITTI Website.

🚀 Usage

To run the Monocular SLAM on KITTI Sequence 00:

# Using the helper script
chmod +x run.sh
./run.sh

Manual execution:

./Examples/Monocular/mono_kitti \
    Vocabulary/ORBvoc.txt \
    Examples/Monocular/KITTI00-02.yaml \
    /path/to/dataset/sequences/00

Controls

Map Viewer: Use the mouse to rotate/zoom the 3D map.

Terminal: Real-time profiling logs are printed to stdout.

Exit: Press Ctrl+C.

📂 Project Structure

src/ORBextractor.cc: Modified class with Grid Distribution logic.

src/OrbCuda.cu: CUDA Kernels and device memory management.

include/OrbCuda.h: Header for C++/CUDA interfacing.

Examples/: Executables for Monocular mode.

🤝 Acknowledgements

This project is based on ORB-SLAM3 by Carlos Campos, Richard Elvira, Juan J. Gómez Rodríguez, José M. M. Montiel and Juan D. Tardós.

Modifications by: Toth Antonio-Roberto

Technical University of Cluj-Napoca