Abstract: Star point detection serves as the core workflow of a star sensor, requiring higher precision detection in complex backgrounds. However, the demand for significant computational resources by high-precision algorithms conflicts with the resource constraints of hardware platforms. To address this issue, this paper proposes a star target detection method based on the YOLOv3-Tiny model, which involves training and optimizing the algorithm using real star observation images. Additionally, techniques such as inter-layer fusion and quantization compression are employed to reduce the computational complexity of the network model and approximately 79.5% of its parameter storage requirements. Furthermore, an efficient convolutional structure design and heterogeneous computing are utilized to maximize hardware resource utilization, enabling the deployment of the algorithm on the Zynq-7020 SoC platform. Experimental results demonstrate that this method achieves over 95% accuracy in star point detection tasks. It also improves upon the limitations of traditional algorithms, which are prone to interference and reduced precision in low signal-to-noise ratio scenarios, significantly enhancing the robustness and adaptability of detection. This approach is well-suited for high-precision star point detection applications.