Point Cloud Recombination: Systematic Real Data Augmentation Using Robotic Targets for Lidar Perception Validation

Abstract

Translate Article

Validating LiDAR-based detection for intelligent mobile systems in open environments requires test data that realistically reflects sensor responses, task-relevant conditions, and environmental context. Virtual simulations allow for controlled scene generation but only approximate sensor-specific effects, such as intensity responses or material-dependent reflections, and require significant effort to construct realistic environments. Real-world data acquired within the intended Operational Design Domain provides genuine sensor realism but offers limited control over influencing factors and lacks situational diversity. We address this gap with a method called Point Cloud Recombination, which systematically augments real outdoor scenes by integrating physical point clouds of target objects acquired in a controlled laboratory setup. Mobile robotic platforms allow systematic variation of design, positioning, and material properties of the targets. A fusion framework combines lab-acquired point clouds with outdoor scenes, while Neural Radiance Field-based mesh generation and subsequent geometric registration ensure realistic integration, preserving phenomena such as point dropouts caused by reflective or absorptive materials. The method is demonstrated by augmenting urban and rural scenes with humanoid targets using an Ouster OS1-128 Rev7 sensor. Evaluation follows Sim2Real principles based on geometric consistency metrics, showing minimal geometrical deviations with an average F1-Hit-Score overlap of 99%. Point Cloud Recombination enables scalable, sensorfaithful test data generation and facilitates systematic, hardwareintegrated perception testing beyond conventional virtual-lidar hardware-in-the-loop approaches, supporting the development of reliable detection functions for safety-critical applications.