We propose an optimization and acceleration solution for Visual-Inertial Navigation Systems (VINS) on UAVs based on heterogeneous computing architecture. As UAV application scenarios diversify, the requirements for autonomous navigation capabilities continue to increase. Visual-Inertial Navigation has become one of the mainstream technologies due to its advantages of low cost, rich information, and high reliability. However, real-time operation of VINS systems on resource-constrained UAV platforms still faces significant challenges, especially in highly dynamic and complex environments. Through a software-hardware co-design approach, this paper addresses the performance bottlenecks of VINS-Mono systems on UAV platforms. The main innovations include: proposing an initialization method combining optical flow and descriptors to improve system robustness in low-texture and fast-motion scenarios; designing an optical flow tracking strategy to replace OpenCV feature extraction, enhancing feature tracking efficiency through hardware acceleration interfaces; developing an adaptive feature selection algorithm to balance computational resources and precision in backend optimization.