RPN+ fast boosted tree: Combining deep neural network with traditional classifier for pedestrian detection

Abstract

The problem of pedestrian detection receives increasing attention due to the rapid development of artificial intelligence technologies. In this paper, we propose a deep neural network based method which combines with a traditional classifier for fast and robust pedestrian detection. Specifically, region proposals generation and feature extraction are implemented using a modified RPN-VGG method. The proposed method is designed to improve system performance on small objects detection. A new classifier, Fast Boosted Tree, is trained based on RPN outputs to obtain the final results. Experiments on Caltech pedestrian dataset demonstrate that the proposed method achieves 8.77% miss rate and has the best known efficiency with state-of-the-art CNN-based detectors. When algorithm efficiency is not considered, detection quality can be further improved to 8.25% miss rate by adding global normalization and optical flow features.