To address the limitations of traditional General Movements assessment in infants, which relies on manual observation and is time-consuming, subjective, and lacks quantitative analysis, a study was conducted on automated assessment methods through infant pose estimation. A pose estimation model named OCEANPose, built upon the VitPose skeleton network, was proposed. By incorporating an Adaptive Occlusion Estimation Weighting module and a cross-attention mechanism, the model improved joint point recognition accuracy, with experimental results showing a 1.2% increase in mean accuracy compared to the baseline model. Furthermore, targeting the high-risk Cramped-Synchronized abnormal movements in GMs, a machine learning-based detection method was developed. This method calculates cosine similarity and autocorrelation values from the joint point sequences extracted by OCEANPose, enabling automatic screening of such abnormal movements based on these two metrics. Experimental results indicate that the proposed approach provides a feasible technical solution for achieving objective and automated GMs assessment.