Computer vision has advanced quickly in recent years and is still spreading to interdisciplinary fields. Nowadays, research on behavioral information about autism has been hampered in the field of autism recognition by challenges with dataset collection, strong privacy concerns that prevent public disclosure, and other issues. In order to deal with these problems and improve the model’s accuracy and efficiency and achieve good results on datasets like Human3.6M, in order to estimate 3D poses from monocular videos, this paper proposes a new BBLMixSTE. It incorporates a DPC-KNN module and combines it with down-sampled data to restore the tokens to their original length. We used this method to de-identify experimental data that we collected from 25 individuals with Typically Developing (TD) and 25 individuals with Autism Spectrum Disorder (ASD). By employing this method, research data for effective behavioral analysis has been provided, and the privacy protection of patients has been ensured in practical application scenarios at the same time. Our model further improved by approximately 4.9% P-MPJPE and 3.5% MPJPE. The experimental results indicate that the approach not only enhances the accuracy of behavioral analysis but also maintains a high standard of privacy due to the careful handling of sensitive information.