Near-Earth Asteroids posed a threat to human civilization, making their monitoring crucial. As the demand for asteroid detection technology increased, precise detection of these celestial bodies became an urgent task to understand their characteristics and assess potential impact risks. To improve asteroid detection accuracy and efficiency, we proposed an advanced image processing method and a deep learning network for automatic asteroid detection. Specifically, we aligned star clusters and overlaid images to exploit asteroid motion rates, transforming them into object-like trajectories and improving the signal-to-noise ratio. This approach created the Asteroid Trajectory Image Data set under various conditions. We modified CenterNet2 network to develop AstroCenterNet by integrating Multi-channel Histogram Truncation for feature enhancement, using the SimAM attention mechanism to expand contextual information and suppress noise, and refining Feature Pyramid Network to improve low-level feature detection. Our results demonstrated a detection accuracy of 98.4%, a recall of 97.6%, a mean Average Precision of 94.01%, a false alarm rate of 1.6%, and a processing speed of approximately 17.86 frames per second, indicating that our method achieves high precision and efficiency.