A Sparse Coding Based Transfer Learning Framework for Pedestrian Detection

Abstract

Pedestrian detection is a fundamental problem in video surveillance and has achieved great progress in recent years. However, training a generic detector performing well in a great variety of scenes has been approved to be very difficult. On the other hand, exhausting manual labeling effort for each specific scene to achieve high accuracy of detection is not acceptable especially for video surveillance applications. In order to alleviate the manual labeling effort without scarifying accuracy of detection, we propose a transfer learning framework to automatically train a scene-specific pedestrian detector starting from a pre-trained generic detector. In our framework, sparse coding is proposed to calculate similarities between source samples and a small set of selected target samples by using the former as dictionary. The similarities are later used to calculate weights of source samples. The weights of initially detected target samples are calculated in a similar way but using the selected target dataset as dictionary. By using these weighted samples during re-training process, our framework can efficiently get a scene-specific pedestrian detector. Our experiments on VIRAT dataset show that our trained scene-specific pedestrian detector performs well and it is comparable with the detector trained on a large number of training samples manually labeled from the target scene.