A New Region Proposal Network for Far-Infrared Pedestrian Detection

Abstract

An automatic region proposal network (ARPN) is proposed to generate bounding boxes with confidence scores for far-infrared (FIR) pedestrian detection. The model consists of two parts: first, the bounding boxes are predicted by the L2 loss function and a module is designed based on a convolutional neural network. This module is simple and only has two layers, each with a 1A-1 kernel. Second, a score map is obtained through FIR pedestrian segmentation based on a feature pyramid network. The scores are taken as the confidence levels for the predicted bounding boxes. To obtain the scores, a new labelling method is also introduced in this paper for FIR image segmentation. We can obtain the bounding boxes per pixel directly and efficiently without any manually designed hyper parameters related to the anchor boxes. To validate the model, this paper uses the LSI, CVC09, CVC14 and SCUT FIR pedestrian detection datasets in the experiments. The datasets consist of different sizes of FIR images collected from several different cameras. The datasets contain different outdoor urban scenes collected at different times, from day to night. The recall vs number of proposals, average recall and recall at IoU 0.5 are used to evaluate the proposed method and log-average miss rate is used to evaluate final detection results. Compared with other algorithms, experiments on most of the data sets also show better performance.