Light-field imaging for distinguishing fake pedestrians using convolutional neural networks

Abstract

Pedestrian detection plays an important role in automatic driving system and intelligent robots, and has made great progress in recent years. Identifying the pedestrians from confused planar objects is a challenging problem in the field of pedestrian recognition. In this article, we focus on the 2D fake pedestrian identification based on light-field (LF) imaging and convolutional neural network (CNN). First, we expand the previous dataset to 1500 samples, which is a mid-size dataset for LF images in all public LF datasets. Second, a joint CNN classification framework is proposed, which uses both RGB image and depth image (extracted from the LF image) as input. This framework can fully mine 2D feature information and depth feature information from corresponding images. The experimental results show that the proposed method is efficient to identify the fake pedestrian in a 2D plane and achieves a recognition accuracy of 97.0%. This work is expected to be used in recognition of 2D fake pedestrian and may help researchers solve other computer vision problems.