Efficient holistic feature basis learning for pedestrian detection

Abstract

Pedestrian detection is an important research area in computer vision and Artificial Intelligence due to its potential applications in pedestrian safety, elderly monitor and care, surveillance, image retrieval and video compression. Many pedestrian detection systems have been proposed and it has been pointed out in state-of-the-art research that feature extraction is one of the significant factors in improving the performance of a pedestrian detector. Therefore, much work has focused on proposing novel feature extraction schemes to improve pedestrian detection. Moreover, most are end-to-end pedestrian detection systems, making it unclear about the contribution of classifiers in the detection pipeline. In this paper, we fill in some of this gap and focus on the classification process and propose feature basis learning for holistic high dimensional feature vectors that are common in pedestrian detection. We experimentally show that it is possible to obtain superior performance by our proposed feature basis learning algorithms even on high dimensional datasets.