Monocular depth estimation on resource-constrained devices faces challenges in balancing accuracy with computational efficiency. This paper proposes an improved U-Net architecture for this task. The key innovation is the integration of an Atrous Spatial Pyramid Pooling (ASPP) module into the network’s bottleneck layer. This integration enhances multi-scale feature perception without substantially increasing computational complexity. To ensure deploy ability on small devices, we incorporate lightweight techniques including depth wise separable convolutions, model pruning, and post-training quantization. Furthermore, a mathematical model is established to convert the network’s relative depth output into absolute physical distance, leveraging camera calibration and reference objects. Experimental results on NYU Depth V2 and KITTI datasets demonstrate that our method achieves competitive depth estimation accuracy. When deployed on embedded platforms like Jetson Nano, the optimized model maintains real-time inference speeds while consuming minimal power. This work provides a practical solution for enabling high-precision depth sensing on mobile and embedded systems.