Hardware-Efficient Local-Expansion-Move Accelerator with Confidence-Driven 3D Label Fusion for Real-Time Stereo Vision

High-precision real-time stereo vision is critical for 3D reconstruction in embedded systems. The local- expansion-move (LE) algorithm shows great potential due to its high precision, but its hardware implementation faces a critical throughput-resource trade-off. To address this, we have improved both the algorithm and hardware design. For the algorithm, we propose WTA-3D label fusion, a new label generation algorithm that combines the original LE with a local-optimal guidance strategy to generate high-confidence labels, reducing iteration dependency and accelerating convergence. For the hardware, using high-level synthesis (HLS), we co-optimize the calculation and cache modules: the calculation module restructures pipelines by decoupling data dependencies between original and new label generation algorithms, while the cache module introduces a compact dual-buffer mechanism to double transfer bandwidth with minimal resource overhead. The optimized accelerator achieves a 3.2× speedup over the original LE design, delivering real-time 50 frames per second (fps) performance with 31% fewer LUT resources and 61% reduced RAM utilization while maintaining state-of-the-art accuracy through lightweight neural network integration.