Abstarct: In this thesis, the identification and tracking of pedestrians in video cameras are intensively and in real time. Some pedestrian tracking applications can be used to identify a particular person or to track down criminals across the city. One of the major challenges of real-time tracking is the need for a lot of memory because the data should be updated continuously. One solution to this is to select the current frxame and a limited number of subsequent frxames.The problem with this method is the lack of sufficient data for training. In such cases, a poor educational model is formed with low data, which, in conditions like brightness changes, variable displacement of the target angle, image blur and blockage are not robust and cause tracing of objects. As long as the education model is dependent on the data and, accordingly, the models are trained, there is a possibility of drift and the wrong tracking. In this thesis, the non-adaptive model and the space structure of features are used for training. Reinforcement learning is a kind of uncontrolled training combined with testing and error and reward methods. Reinforcement learning always uses multiple paths to identify and track targets, and is obstructive against intelligence, and can recover lost frxames and re-target tracing. In this thesis two stages of compression on raw data and compression of thinner data have been used. Initial compression reduces computational complexity and, as a result, saves memory usage. For thrust to reach the minimum, compression on thinning data is used, which always selects the most suitable data for tracking, and does not consider the least important features. And makes it possible to provide a strong and accurate model for training. The results show that the proposed algorithm has high efficiency against noise and severe obstruction, so that the accuracy of tracking on the databaxse was 71.5% and 36.7%, respectively.