3D Pedestrian Trajectory Prediction using Deep Learning from Kinect Data
|
Akbar Jafari , Ali Hosseininaveh * , Mojtaba Mahmoodian  |
K.N.Toosi University of Technology |
|
Abstract: (1318 Views) |
Pedestrian trajectory prediction is a critical challenge in the fields of computer vision and intelligent transportation systems, as it directly impacts the safety and decision-making capabilities of autonomous systems. Most existing approaches rely on two-dimensional (RGB) data and recurrent neural networks such as LSTM (Long Short Term Memory), which neglect the depth dimension and therefore fail to accurately estimate distances between pedestrians and surrounding objects. In this study, we propose a 3D-LSTM (Three Dimension LSTM) model that utilizes RGB-D data obtained from a fixed Kinect sensor to predict pedestrian positions in metric three-dimensional space. The proposed framework includes depth extraction from stereo images, coordinate normalization, and LSTM-based sequence modeling to forecast future pedestrian positions in the (X, Y, Z) coordinates. Experimental evaluations conducted on the École Polytechnique Fédérale de Lausanne (EPFL) dataset demonstrate that the 3D prediction accuracy (average RMSE: 15.7 cm) is comparable to conventional two-dimensional methods while additionally providing real-world distance and spatial interaction information that is crucial for collision avoidance and motion planning. The results indicate that incorporating the third dimension does not degrade performance; instead, it enhances the ability of intelligent systems to make safer and more informed decisions in dynamic environments. This approach lays the groundwork for advanced navigation and autonomous driving systems with enhanced three-dimensional situational awareness.
|
|
Keywords: Pedestrian Trajectory, Trajectory prediction, deep learning, LSTM Network |
|
Full-Text [PDF 1105 kb]
(25 Downloads)
|
Type of Study: Research |
Subject:
Aerial Photogrammetry Received: 2023/06/10 | Accepted: 2024/05/26 | ePublished ahead of print: 2025/03/17 | Published: 2025/08/31
|
|
|
|
|
Send email to the article author |
|