Pupil localization plays a crucial role in human-computer interaction and biomedical computing applications. Currently, many sophisticated pupil localization algorithms are designed to detect and locate the pupil position using one single image. However, pupil movement is a continuous process. Therefore, when the pupil position cannot be accurately detected and located in one current frame, the pupil position can be inferred by combining information of previous frames. This approach can more effectively handle difficult and challenging situations such as reflections, pupil occluded by eyelashes and blinks, as well as off-center pupil positions and motion blur. Consequently, it can significantly improve the accuracy and robustness of pupil detection, decreasing localization errors. To address these challenges, propose a pupil detection algorithm based on deep learning using multiple consecutive images. This algorithm enhances the standard Unet encoder-decoder structure by incorporating multi-frame information from continuous eye tracking scenes for improved pupil detection. By combining convolutional neural networks with convolutional long short-term memory networks and a convolutional block attention module, we introduce a hybrid semantic segmentation network. Experiments on a large-scale dataset demonstrate that the proposed method outperforms existing pupil detection algorithms, achieving a mean intersection over union score of 96.78% and a root mean square error value of 3.83, especially in challenging situations.