Edge-Accelerated Monocular Depth Estimation: A Quantized U-Net on Kria KV260 for Real-Time Interior Scenarios

Real-time monocular depth estimation (MDE) is pivotal for edge applications like autonomous indoor navigation and AR/VR, yet achieving it on power-constrained devices remains a significant challenge. This work presents an optimized, INT8-quantized U-Net model specifically designed for the Xilinx Kria KV260 FPGA. Trained on the NYUDepthV2 dataset and compiled using the Vitis AI toolchain, our implementation achieves a throughput of 41.03 FPS with a latency of 24.3 ms and an RMSE of 0.9553 m. This represents a 2.27× speedup over a desktop-grade Intel i7-11700K CPU while maintaining comparable precision. Our contributions include: (i) a FPGAtailored U-Net (ii) an efficient quantization-to-DPU workflow, and (iii) a demonstration of the real-time MDE indoors on the KV260.