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).

Screenshots:

🤝 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