Design and implementation of object detection system based on Zynq hardware acceleration

Zynq integrates programmable logic (PL) and processing system (PS), offering an efficient and flexible solution for the deployment of deep learning models on embedded devices. The platform's parallel computing capability and hardware acceleration features facilitate real-time inference in resource-constrained environments. This paper investigates the advantages and applications of deploying an enhanced YOLOv3 algorithm on the Xilinx Zynq UltraScale+™ MPSoC AXU5EV-B platform, analyzing power consumption, hardware resources, and performance requirements. The optimized network model employs depth wise separable convolutions to substitute original convolutions, and compares the variations in parameter count before and after optimizing the YOLOv3-Tiny model. Through software and hardware co-design, integrating the optimized YOLOv3 algorithm with Zynq's hardware acceleration considerably enhances the operational efficiency and real-time performance of the object detection system. Experimental results indicate that the enhanced YOLOv3 object detection system based on the Zynq platform not only significantly improves the speed of deep learning inference but also markedly reduces power consumption compared to other hardware platforms, fulfilling real-time detection requirements in complex scenarios.