The elderly have been the subject of significant concern in recent years due to the growth of their population, the reduction of caregivers' population, the growing cost of long-term medical care, and the people's preference for having an independent life. Recent statistics show that the proportion of the world's population aged over 60 years will increase from 12% in 2015 to 22% in 2050. Therefore, it is a requirement to develop intelligent systems that allow improving the quality of people's lives, especially dependent people, in the context of Ambient Intelligence (AmI). AmI-based environments, e.g., Ambient Assisted Living (AAL) systems, are intended to provide context-aware services to the users. Designing AAL systems that can automatically recognize human activities, human behaviors, abnormal human behaviors, and provide well-being recommendation system poses several challenges. The latter are addressed in the literature without considering sufficiently user's context, which lead to erroneous and inconsistent human activity and behavior recognition; i.e., exploiting user's context, such as relationships between activity predictions prevents classification errors and inconsistencies. In this thesis, different frameworks are proposed to provide a more comprehensive description of user's activities and behaviors by exploiting his/her context and the advantages of hybrid approaches that combine data-driven and knowledge-driven approaches. Four datasets are used to evaluate the performance of the proposed frameworks: the Opportunity, the Human Movement and Ergonomics, the Orange4Home, and the UCI HAR datasets. Two multi-label human activity recognition frameworks are firstly proposed to deal with erroneous and inconsistent predictions. The first framework exploiting machine-learning models and Bayesian networks is proposed to detect and correct activity classification errors. In the second framework, machine-learning models, ontological reasoning, and Bayesian reasoning are used to detect and correct inconsistent predictions. The third framework, based on machine-learning models, ontological representation of human behavior context, and commonsense reasoning, is proposed for human behavior recognition. In the fourth framework, the recognized human behaviors and probabilistic commonsense reasoning are used to recognize abnormal human behaviors. Lastly, a context-aware adaptive system exploiting the recognized human behaviors and probabilistic commonsense reasoning is proposed to provide adaptive well-being recommendations based on the user's preferences.