As a high-precision attitude measurement device, star sensors are widely used to determine the attitude of spacecraft in space. With advancements in detection technology, all-day star sensors have been applied to atmospheric platforms such as aircraft, vehicles, and ships. However, their performance is constrained by the low signal-to-noise ratio (SNR) of faint stars under daylight backgrounds. Traditional single-frame threshold segmentation struggles to extract faint stars, while existing multi-frame adding alignment methods mostly rely on external gyroscope attitude information, which is affected by gyroscope drift errors and is unsuitable for long-duration adding. Additionally, inter-frame image transformation models often use approximate models, introducing systematic errors that lead to insufficient alignment accuracy under large attitude changes, resulting in limited SNR improvement. This paper proposes a multi-frame adding method based on inter-image correlation, establishing an unbiased inter-frame image transformation model to optimize alignment accuracy and significantly enhance the extraction capability of faint stars.This method initially extracts a small number of star points from a single frame and then utilizes prior information to rapidly extract visible star points from adjacent frames using a tracking gate. It establishes inter-frame correlations and calculates the relative attitude changes of the star tracker between frames based on the star vector information of correlated star points. The method then reprojects these points onto the current frame's image plane based on this relative attitude, achieving unbiased transformation of inter-frame star images. These star images are then streamed and recursively added to enhance the signal-to-noise ratio (SNR) of the star points. Finally, threshold segmentation is used for coarse extraction, and Gaussian surface fitting is employed for precise localization.The simulation analysis demonstrates that the alignment error of the traditional rigid-body transformation and affine transformation models generally increases with the augmentation of the attitude transformation angle when superimposing faint stellar points (Fig.6). Under a 3° attitude variation, the systematic error of the traditional rigid-body transformation and affine transformation models exceeds 5 pixels across a large area (Fig.5). This error prevents the energy of faint stellar points from concentrating, thereby hindering the effectiveness of the superimposition enhancement. In contrast, the attitude projection model proposed in this study does not exhibit such systematic errors. The sequence star image superposition experiments reveal that the proposed method can effectively extract faint stellar points that are invisible in a single frame. The superimposed faint stellar points exhibit complete morphology without any trailing effects (Fig.9), and their central positions align with the theoretical positions, indicating that the method can avoid alignment errors during the superposition process of faint stellar points in image transformation. Compared to the traditional rigid-body transformation and affine transformation models, the proposed method extracts a greater number of faint stellar points, demonstrates enhanced detection capability, and achieves signal-to-noise ratio improvements of 85.0% and 118.5%, respectively. Additionally, the positioning errors are reduced by 89.4% and 77.9%, respectively.This paper addresses the issue of stellar point alignment errors in traditional multi-frame superposition methods for all-day star sensors under dynamic conditions by proposing a novel multi-frame superposition method based on image autocorrelation. By establishing precise correspondence relationships of stellar points between adjacent frames, the relative attitude changes of the star sensor are determined, and an unbiased transformation of the star map is achieved based on the spatial projection model of the camera. This approach fundamentally eliminates the systematic biases introduced by traditional approximate models such as rigid-body transformation and affine transformation. The proposed method significantly reduces alignment errors during transformation, effectively enhances the signal-to-noise ratio of stellar points, improves the capability to extract faint stellar points, and markedly increases positioning accuracy. This method effectively resolves the application limitations caused by model errors and external dependencies in traditional multi-frame accumulation techniques, providing both theoretical support and engineering practice solutions for the autonomous attitude determination capability of all-day star sensors in highly dynamic environments.