Pedestrian Activity Classification With Hybrid Deep Learning Models Using Elevated LiDAR 3-D Point Clouds

Abstract

Recent advances in light detection and ranging (LiDAR) and artificial intelligence have enabled new opportunities for traffic surveillance and pedestrian monitoring in urban environments. However, most existing systems focus on object detection and tracking, while reliable pedestrian activity recognition under realistic degradation remains insufficiently addressed. This article presents an elevated LiDAR-based framework that captures overhead 3-D point clouds for object recognition and multiclass pedestrian activity classification. Despite progress in 3-D perception, three major challenges persist: 1) the scarcity of activity-labeled real-world LiDAR datasets suitable for supervised learning; 2) limited robustness of deep learning models when LiDAR signals are affected by noise, sparsity, and occlusion; and 3) the lack of unified pipelines that jointly integrate 3-D detection with spatial and temporal behavior analysis. To address these challenges, we develop a Blender-based elevated LiDAR dataset simulating diverse urban scenarios and extend the KITTI tracking dataset with activity annotations for real-world validation. The proposed framework combines PV-RCNN for 3-D pedestrian detection with a hybrid PointNet–long short-term memory (LSTM) architecture for spatial–temporal activity recognition, guided by a dynamic weighting fusion mechanism. Experimental results show that the proposed model outperforms PointNet, LSTM, and transformer-based baseline across all the activity classes, achieving 84.7% accuracy on synthetic data. Under severe LiDAR corruption, performance degrades by only 3.4 percentage points. On the KITTI-based dataset, the framework achieves 73.5% accuracy, surpassing single-stream baselines by 6.5–23.0 percentage points. These results demonstrate that joint spatial–temporal reasoning yields improved comparative robustness under controlled LiDAR degradations while maintaining consistent cross-domain generalization under real-world conditions for pedestrian activity recognition in smart-city environments.

Department(s)

Electrical and Computer Engineering

Comments

National Science Foundation, Grant 2052528

Keywords and Phrases

3-D point clouds; elevated light detection and ranging (LiDAR); gait analysis; human activity recognition (HAR); noise-robustness models; pedestrian monitoring; real-world data testing

International Standard Serial Number (ISSN)

1558-1748; 1530-437X

Document Type

Article - Journal

Document Version

Citation

File Type

text

Language(s)

English

Rights

© 2026 Institute of Electrical and Electronics Engineers, All rights reserved.

Publication Date

01 Jul 2026

Link to Full Text

Share

 
COinS