Human Activity Recognition (HAR) have become an important part to some clinical decision support systems that provides vital contextual information that enhances the monitoring for self-management of chronic conditions such as Musculoskeletal Disorders, chronic pain, and increases the effectiveness of various mobile health. To make such services available to a broader population, one should use wearable devices that most users already have, such as smartphones and smartwatches, which able to provide information about user activities through the analysis of sensor data. HAR is typically modelled as a classification task where sensor data associated with activity labels are used to train a classifier to recognise automatically the future occurrences of these activities. Generally the performance of a generic HAR model trained from a general population (subject-independent) significantly decreases when it is tested on a specific user, due to inter-subject variability like variations in activity patterns, behavioral status of users, gait or posture between different users. To solve this problem, we propose a auto-supervision formalism based on the theory of a Sequential Monte Carlo (SMC) filter to automatically build a personalized HAR classifier. The suggested approach uses different components based on the SMC filter steps to automatically and iteratively approximate the target distribution as a set of temporal samples in order to personalize the HAR model towards a target user. Moreover, we put forward a likelihood function that combines temporal information extracted from the target user, to favor the selection of target samples associated with the right label. Our experiments showed that in general, personalization is critical when an HAR system is used for a new user. The experimental results show that our proposed framework improves the accuracy of HAR on a new user by 50% on average compared to the case of using a model for a new user with no personalization on several public HAR datasets.