The invention relates to a quick real-time discrimination type tracing method based on multi-local-feature learning. The method comprises the following steps of: single-frame decomposition is carried out on a video, a target to be traced is marked in an initial frame through combination with early-stage action identification work or an artificial marking method, then, candidate blocks are obtained in a candidate area through dense sampling, and the local feature of each candidate block is independently calculated; then, on the basis of a circulant matrix and a relevant filter, a classifier is trained in a Fourier domain, a detection formula is used for calculating correlation between each candidate block in a current frame and the target also in the Fourier domain in a position detection stage, and an area with the highest correlation is selected as a prediction position of the target of the current frame; and finally, in the prediction position, a position detection result which has a smallest difference with a previous frame of target is selected as a final tracing result of the current frame, and a new target feature is used for training a new classifier. Except the initial frame, the tracing of other frames is characterized in that position detection is firstly carried out, and then, the training of a new regression function is carried out.