A pedestrian classification method based on transfer learning

Pedestrian detection is a challenging research task of computer vision, which can be seen as a classification problem in the sliding window framework. Many supervised learning based methods require a large number of labeled data for training. However, training and testing data are not independent identically distributed in most cases, due to the complex background, and it is expensive to re-collect and label the data. This paper proposes a semi-supervised method for pedestrian classification, which is based on transfer learning and sparse coding and just requires a small quantity of labeled data. Firstly, we use sparse coding to learn a slightly higher-level, more succinct feature representation from the unlabeled data that randomly downloaded from the Internet. Then we apply this representation to the target classification problem by transfer learning. The quantitative experiment results demonstrate that this method can improve the performance of pedestrian classification and just needs only a few labeled data.