Limited processing resources, energy restrictions, and latency sensitivity on mobile devices provide special difficulties for object detection. This study evaluates the deployment of YOLOv5s on iOS devices and proposes a multi-stage optimization framework combining structural pruning and hybrid post-training quantization. The optimized model achieves 85.4% mean average precision (mAP) and sustains 35 frames per second (FPS) on an A14 Bionic-powered device, while reducing model size by 73% and improving energy efficiency by 15% compared to the baseline. The approach introduces a dynamic quantization scheme and a redundancy-aware pruning protocol specifically tailored for mobile hardware. Empirical evaluations across real-world scenarios-including AR environments and robotic systems, validate the effectiveness of the optimizations. This work presents the feasibility of advanced vision systems on limited edge devices and offers a repeatable, low-latency deployment pipeline for mobile object detection.