Abstract
Real-time image processing on low cost embedded systems is still a challenging research area. For this embedded platform, there is a trade-off between accuracy and processing time. We proposed a pedestrian detection method for thermal images that can perform in real-time on a Raspberry Pi embedded system while still keeping the accuracy high. Our detection framework is based on the conventional HOG-based pedestrian detection method in which the processing time is improved by performing the HOG computation only on the regions of interest. These regions of interest are obtained using foreground segmentation where we compute the segmentation process on the half-size image to further reduce the processing time. In the experiment, our two proposed methods can significantly reduce the processing time while still have higher precision and recall compared to the two baseline approaches.