Factorized Spatio-Temporal Convolutions for Human Pose Estimation from Planar Lidar

2026-07-23

AI Summary

This work addresses the challenge of efficient human detection and relative 2D pose estimation for resource-constrained service robots equipped only with a surround-view planar LiDAR. The authors propose a lightweight network featuring a spatio-temporal separable convolution architecture that separately models spatial features from individual LiDAR scan lines and temporal dynamics across frames—a novel application of such an architecture to pose estimation from planar LiDAR sequences. To circumvent the need for manual annotations, cross-modal self-supervision is established using overlapping regions with RGB-D sensors. Experimental results demonstrate that, compared to baseline methods with similar model size, the proposed approach reduces distance, positional, and orientation errors by 38%, 28%, and 15%, respectively, while enabling real-time inference on CPU. Its effectiveness is validated on both the FROG dataset and real-world scenarios.

Abstract

Localizing nearby humans and estimating their facing direction are key capabilities for safe navigation and socially aware human-robot interaction. Many pose-estimation pipelines target cameras and 3D LiDAR or assume GPU-class compute, whereas service robots are often equipped only with omnidirectional planar LiDARs and modest onboard processors. We address omnidirectional human detection and relative 2D pose estimation from planar LiDAR sequences with a lightweight network based on Space-Time Blocks, which explicitly separate spatial processing along scan rays from temporal aggregation across scans. Our network processes 360° LiDAR sequences to output per-ray human presence, distance, and relative orientation. We train it via cross-modal self-supervision from a narrow RGB-D body tracker in the sensors' overlap region, removing the need for manual LiDAR labels. Quantitative experiments show that our approach consistently outperforms a parameter-matched baseline model, reducing errors in distance (-38%), position (-28%), and orientation (-15%). We further benchmark on the public FROG dataset, report real-time CPU inference on a service robot, and validate with in-field demonstrations, supporting its suitability for spatial perception on computationally constrained service robots.

Problem

• human pose estimation
• planar LiDAR
• service robots
• spatio-temporal perception
• computationally constrained systems

Innovation

• Factorized Spatio-Temporal Convolutions
• Planar LiDAR
• Self-supervised Learning
• Human Pose Estimation
• Lightweight Neural Network