Human Object Identification for Human-Robot Interaction by Using Fast R-CNN

摘要

This paper proposes a human object identification by using a simplified fast region-based convolutional network (R-CNN). Human identification is a problem of considerable practical interest. Here, we propose the state-of-the art method which is tested for major pedestrian datasets. Human detection consists of the body part detectors which detect head and shoulder, torso, and pair of legs, with three, two and four different appearances respectively. These detectors are integrated as to identify the human object with different poses. Fast R-CNN is a well-known method for object recognition using deep CNN. Hybrid body part detector demonstrates the merits for partially occluded human detection by integrating the scores of the individual part detectors based on the occlusion map. The highest merging score is the best configuration to evaluate the detection score of the human detector. Experiments on two public datasets (INRIA and Caltech) show the effectiveness of the proposed approach.