Fast Person Detector with Efficient Multi-level Contextual Block for Supporting Assistive Robot

Abstract

The robotic demand a vision method to work in real-time on embedded devices. Besides, an assistive robot requires person detection, which is widely used to help automatically interact with the user. This work presents a fast real-time person detection (Fast-PdNet) to localize human areas implemented on a Jetson Nano. This device has been commonly used as an embedded system and is suitable for synchronizing sensors and actuators. The proposed architecture contains layers of Convolutional Neural Network consisting of two main modules: backbone and detection. An efficient extractor module with a multi-level contextual block is employed to extract the spatial features quickly. It avoids high-cost computing to distinguish interest features of the human body and background features. The lightweight learning attention selects suspected specific features area without generating excessive parameters. The end-to-end training was conducted on MS COCO 2017 to generate efficiently weighted models. The Fast-PdNet achieves competitive performance with other light detectors evaluated on the MS COCO 2017, PASCAL VOC 2007, and 2012 datasets. Moreover, this detector can run 35 frames per second when working in real-time on Jetson Nano.