Design and Research of AI Accelerators for Onboard Computing Platforms Targeting YOLOs Network

Convolutional Neural Networks (CNNs) play a crucial role in computer vision, with wide applications in image recognition, object tracking, and trajectory optimization. Due to constraints in computation, power consumption, and size in aerospace vehicles, deploying CNN algorithms on target platforms presents a significant challenge in aerospace applications. This paper proposes an accelerator design for the YOLOv3-tiny network tailored to aerospace applications. By employing techniques such as loop partitioning, loop unrolling, and weight sharing, the proposed accelerator enhances computational performance and data reuse efficiency. Additionally, by integrating convolution, activation, and pooling operations into a unified module, this approach reduces the need for intermediate buffers, saving memory and improving acceleration performance. Compared to accelerators that require buffering of convolution results, the proposed design demonstrates superior performance. The accelerator is implemented on the ZYNQ7020 FPGA platform, achieving a peak computational performance of 114.53 GOPS.