Real-Time Stairs Detection and Terrain Classification: Evaluating LiDAR Sensor Performance

Abstract

The growing use of wearable and autonomous systems in dynamic environments requires reliable terrain classification and obstacle detection to ensure user safety, adaptability to environmental conditions, and optimize system performance. This thesis aims to develop a complete real-time processing pipeline for terrain classification and stairs detection using LiDAR sensors, focusing on the comparative performance of both low- and high-accuracy sensors: the Cyglidar D1 and RoboSense RS-LiDAR-M1 v2.

The research introduces a novel end-to-end pipeline that integrates innovative feature extraction methods with efficient classification algorithms, optimized for real-time processing of 3D point clouds. The study utilizes Support Vector Machines (SVM) with linear and non-linear kernels to classify terrain or stairs, deploying a combination of wearable and tripod-mounted configurations to collect data. These models were implemented on the Jetson Nano, an embedded platform, to evaluate real-time feasibility. Key techniques included normal vector extraction as part of the feature extraction process, noise simulation to account for lower-cost sensor conditions, and comprehensive real-time performance analysis, enabling robust classification across diverse scenarios.

The results indicated that while the Cyglidar D1 sensor achieved a commendable F1 score of 0.96 in stairs detection, its performance in terrain classification was hindered by noise and overlapping features, particularly between plain and grassy terrains. On the other hand, the RoboSense RS-LiDAR-M1 v2 excelled, achieving an F1 score of 0.99 in terrain classification with minimal processing delay. The system proved robust against noise, maintaining an F1 score of 0.99 even with 10 mm of added Gaussian noise.

These findings validate the effectiveness of the proposed pipeline, demonstrating its ability to balance sensor cost, noise tolerance, and algorithmic efficiency while delivering robust performance in real-time applications. The research provides valuable insights into the trade-offs between sensor accuracy and cost, contributing to advancements in LiDAR-based terrain classification and obstacle detection systems. Moreover, it underscores the importance of preprocessing and sensor selection in enhancing classification reliability in real-world scenarios.