The invention discloses a pedestrian detection method based on self-learning. The method comprises the following specific steps that: firstly, training an AdaBoost-based cascade classifier as an off-line classifier, meanwhile, using one group of public pedestrian photos to train a Gaussian mixture model, and adopting HOG (Histogram of Oriented Gradient) feature and position information for feature coding; then, adopting the off-line classifier of a low threshold value to carry out pedestrian detection on a specific scene, and outputting the confidence score of a candidate object; then, picking up a high confidence score as a positive sample and a low confidence score as a negative sample, and using a Gaussian mixture model to show the candidate detection object again; and finally, using a SVM (Support Vector Machine) classifier to train a pedestrian classifier with discriminating ability on line, predicting the candidate object again, and estimating an output probability. By use of the method, the problem that a traditional pedestrian detection method can not carry out adaptation on a specific scene is solved, and the method has a certain promoting effect on a pedestrian detection technology under the specific scene. Compared with a traditional pedestrian detection method, the pedestrian detection is characterized in that a recognition rate is obviously improved.