A self-learning framework for pedestrian detection

The performance of offline trained pedestrian detectors significantly drops when they are applied to the specific scene. Although manual labeling can improve detection performance, it requires a lot of human effort. In this paper, a self-learning framework is proposed for pedestrian detection, which can adapt any offline trained detector to a specific scene and obtain a better performance. Firstly, Cascade classifier is used as an offline classifier, while a Gaussian Mixture Model (GMM) is trained using a set of public pedestrian photos. Next, a low threshold offline classifier is used to perform pedestrian detection on a specific scene and the confidence score of candidate detections is obtained. Then, samples with high confidence scores are selected as positive samples, while those with low confidence scores are taken as a negative samples, and GMM is used to represent the candidate detection again. Finally, a discriminative pedestrian classifier is trained online using the SVM classifier to re-estimate candidate objects. Experimental results on public and self-made datasets show that the proposed method can improve the accuracy of the generic pedestrian detector and significantly outperforms the traditional methods.