Vision-based human motion analysis has received huge attention from researchers because of the number of applications, such as automated surveillance, video indexing, human machine interaction, traffic monitoring, and vehicle navigation. However, it contains several open problems. To date, despite very promising proposed approaches, no explicit solution has been found to solve these open problems efficiently. In this regard, this thesis presents a model-based human upper body pose estimation and tracking system using interest points (IPs) in real-time video. In the first stage, we propose a novel IP-based background-subtraction algorithm to segment the foreground IPs of each frame from the background ones. Afterwards, the foreground IPs of any two consecutive frames are matched to each other using a dynamic hybrid localspatial IP matching algorithm, proposed in this research. The IP matching algorithm starts by using the local feature descriptors of the IPs to find an initial set of possible matches. Then two filtering steps are applied to the results to increase the precision by deleting the mismatched pairs. To improve the recall, a spatial matching process is applied to the remaining unmatched points. Finally, a two-stage hierarchical-global model-based pose estimation and tracking algorithm based on Particle Swarm Optimiation (PSO) is proposed to track the human upper body through consecutive frames. Given the pose and the foreground IPs in the previous frame and the matched points in the current frame, the proposed PSO-based pose estimation and tracking algorithm estimates the current pose hierarchically by minimizing the discrepancy between the hypothesized pose and the real matched observed points in the first stage. Then a global PSO is applied to the pose estimated by the first stage to do a consistency check and pose refinement.