<p>Although deep-learning-based visual detection has advanced fall monitoring, conventional fall detection models still exhibit high computational complexity and slow inference speed, which restrict their deployment on resource-constrained embedded platforms. To address these challenges, this study proposes a lightweight fall detection algorithm based on MobileNetV2 and its efficient hardware implementation on a Zynq SoC. Knowledge distillation is employed, where an ECA-enhanced ResNet50 serves as the teacher model and a lightweight MobileNetV2 acts as the student model. In addition, batch-normalization fusion and quantization-aware training further improve accuracy while compressing model size. For hardware acceleration, a convolution-fusion block, pipelined architecture, and dual-buffer optimization are designed to enhance computational efficiency and data throughput. Experimental results demonstrate that the optimized MobileNetV2 achieves an accuracy of 96.3%. After compression, the model size is reduced to one-fourth while maintaining 95.6% accuracy. The Zynq implementation reaches 16.32 FPS with a power consumption of 2.765 W, achieving a 16× speedup compared with an Intel i5-10400H processor. These results verify that the proposed framework provides an effective solution for real-time and low-power fall detection on embedded platforms.</p>