Gao, Zhenyu (Institut des systèmes intelligents et robotique (ISIR) - Sorbonn), Wang, Ze (Institut des systèmes intelligents et robotique (ISIR) - Sorbonn), Saint-Bauzel, Ludovic (Sorbonne Université), BEN AMAR, Faiz (Institut des systèmes intelligents et robotique (ISIR) - Sorbonn)

Human Orientation Estimation from 2D Point Clouds Using Deep Neural Networks in Robotic Following Tasks

Scheduled for presentation during the Regular Session "AI-based Robot Control II" (WeBT2), Wednesday, July 16, 2025, 14:40−15:00, Room 106

Joint 10th IFAC Symposium on Mechatronic Systems and 14th Symposium on Robotics, July 15-18, 2025, Paris, France

This information is tentative and subject to change. Compiled on July 16, 2025

Abstract

Human orientation is a crucial information for a mobile robot that has to follow a person moving freely. This paper introduces a deep neural network architecture using 2D point clouds from a LiDAR mounted at knee height on a mobile robot platform named SUMMIT-XL to predict human orientation. The proposed architecture integrates Principal Component Analysis (PCA) to reduce input data dimensionality and enhance neural network generalization. It combines a Convolutional Neural Network (CNN) for feature extraction and a Recurrent Neural Network (RNN) with a time windower to capture implicit gait dynamics within a time-series point cloud. This architecture was evaluated on a dataset collected over six hours from five volunteers, comprising point cloud data and ground truth of human orientation during complex walking patterns. The general model, trained on a combination of multi-person data, achieved a Mean Absolute Error (MAE) of less than 13 degrees. Meanwhile, the customized model, trained on individual data, achieved an MAE of less than 7 degrees and showed no delay.