The invention discloses a humanoid robot expression recognition method based on a dense convolutional neural network. The method comprises the steps of positioning a face position in a single frame image according to a directional gradient histogram feature and a support vector machine algorithm, and matching the face reference points of the face image by using the regression tree set, correctingand recording the face image according to a matching result, identifying the corrected face image by using the dense convolutional neural network, obtaining expression classification according to an identification result, and transmitting an expression classification result to the humanoid robot. The technical scheme provided by the invention is based on face positioning and face correction. The dense convolutional neural network is used to combine the flow of face feature extraction and expression classification so as to realize the expression recognition of the face, and finally the classification result is provided for the humanoid robot in real time, so that the effect of human-computer interaction is achieved, the accuracy is ensured, the real-time performance is high, and the safetyof pedestrians is effectively ensured.