The difficulty of person re-identification is that the same person images in different cameras are significantly different, which is difficult to stably describe the images by a single feature, while the fusion by a variety of features can’t distribute their weights exactly. To solve the problems, a person re-identification algorithm based on multi-kernel support vector machine by multi-instance learning is proposed. Firstly, the blocked color features in HSV space and local features of SIFT from the same people image under different cameras are extracted, and the bag of words are constructed to SIFT features. Both of them are taken as two instances and encapsulated as a bag especially. Secondly, the multi-kernel support vector machine model is optimized, the bags are trained by the linear fusion kernel between Gaussian and polynomial, and then the optimal weighting ratio is obtained by multi-instance learning. Finally, this algorithm is tested on the VIPeR dataset, the accuracy rate of recognition is an average accuracy of ten times experiments, and expressed by CMC curves. At the same time, the matching result of the sample is also sorted. The experiments show that the robustness and recognition rate of this algorithm achieve the same and even better results while compared with several state of the art algorithms.