Abstract: The tracking of objects in real time video can significantly facilitates works of many people, such as, for example, security guards or road policemen. This paper presents algorithm for tracking pedestrians as some objects in video frames. Pedestrians are only one example the using of algorithm, but proposed algorithm can be useful in many civil and military areas. We use the principle of repeatedly identification of objects, which includes algorithms FAST and HOG for searching object in frame by key points. It is shown in work the advantage of developed algorithm compared to methods and algorithms with same or similar results. Our algorithm has high performance and can be used in low power built-in real time systems. In spite that developed algorithm in accuracy is below methods of deep learning, it is not require big data base of pedestrians and vast amount of processing power for training of neural nets. Algorithm presented in this paper have higher speed in comparison with deep learning methods and other methods, proposed earlier by several authors. We show some deficiencies of algorithm, its correction in future will significant increase the accuracy of method.
Index terms: computer vision, tracking, pedestrian, histogram of oriented gradients, FAST, Re-Identification.