Validation and accuracy estimation of a novel measurement system based on a mobile robot for human detection in indoor environment

This paper validates an indoor measurement system for human detection using a RGB camera installed on a mobile robot at three different robot’s head configuration and in two suboptimal collected scenarios. Images are processed with three algorithms for human detection i.e., Histogram of Oriented Gradients (HOG), Viola-Jones Haar Cascade Classifier (Haar cascade classifier), You Only Look Once (YOLO-v3) trained with COCO dataset and their accuracies in detecting humans are computed. For these algorithms, dependences from the robot head configuration and from the robot-subject distances are assessed in the two suboptimal scenarios. Results show that in the first suboptimal scenario of the proposed measurement system, YOLO-v3 algorithm provides the best accuracy value in detecting humans (99,9%) while the best accuracy results for the two other algorithms (65,7% for HOG and 70,8% for Haar cascade) are reached for a robot head configuration of 40°, independently from the robot-subject distances. The second suboptimal scenario is performed to detect conditions in which YOLO-v3 fails. Results indicate that in two user configurations YOLO-v3 fails that could be attributed to the COCO dataset with which the YOLO-v3 algorithm is trained and to the proposed suboptimal test conditions.