Abstarct: Nowadays, development of smart homes, aiming at human convenient lives, has garnered substantial attentions. This issue is motivated by assisting elderly and infirm individuals, providing high levels of security, energy management, and medical healthcare. Making decisions on how to provide essential services, is carried out baxsed on the residents’ activities. So, human activity recognition (HAR) is of crucial importance in such environments. For this purpose, various sorts of sensors are hired to collect users’ data. The acquired data is further processed via data mining and machine learning techniques to infer human activities. Most of the current HAR approaches are baxsed on simplified presumptions and real world complex circumstances are seldom taken into account. Such presumptions are restrictive and not applicable in many realistic scenarios. Some of the predominant common presumptions include: considering sequential activities, single subject environments, and neglecting the uncertainty in sensors’ data and inference. It is necessary to reduce the effects of these assumptions in HAR schemes to make them applicable in empirical situations. In the presented thesis, we aim at developing a HAR scheme in such circumstances. Regarding the above-mentioned circumstances, we propose a three laxyered HAR architecture. The laxyers pertain to data acquisition, single-subject HAR, and multi-subject HAR, respectively. In the first laxyer, namely data acquisition, users’ data are collected by various environmental sensors and catered for the next laxyers as a sensor data stream. The second laxyer, i.e. single-subject HAR laxyer, achieves HAR regarding the mitigation of uncertainty in inference. In the third laxyer, i.e. multi-subject HAR laxyer, individual data traces are modeled and extracted out of users’ aggregate data. Afterwards, users’ activities are inferred on the extracted data traces using the procedures presented in the second laxyer. Results show that the proposed architecture can outperform common HAR approaches that are generally baxsed on probabilistic graphical models.