Title: Virtual data generation method for simulation scenes based on scene flow prediction

Authors: Lizhen Jiang; Xu Lan

Addresses: School of Media and Art Design, Guilin University of Aerospace Technology, Guilin, Guangxi, China ' School of Art & Design, Guilin University of Electronic Technology, Guilin, Guangxi, China

Abstract: A virtual data generation method based on scene flow prediction is proposed to address the problem of high false positives and inability to ensure feature consistency after virtual data encoding in simulation scenarios. Firstly, a variational auto-encoder is introduced to determine the relationship between virtual data in different dimensions through back-propagation performance; then, with the help of deep semantic information, automatic annotation of simulation scene data is completed; finally, different displacement changes in the three-dimensional space of the data object are determined to obtain dense scene streams. Based on this, the scene stream prediction method is used to encode the entire virtual data and fuse it. Combined with the matching of encoded features, the consistency of the encoded virtual data features is ensured. Finally, the objective function is introduced, and a data generator is set up to achieve virtual data generation in simulation scenarios. The results show that the proposed method reduces the false positive rate of virtual data generation in simulation scenarios and has good performance.

Keywords: scene flow; prediction; simulation scenario; virtual data; generation method; deep semantic information; back-propagation.

DOI: 10.1504/IJCAT.2024.143295

International Journal of Computer Applications in Technology, 2024 Vol.74 No.4, pp.351 - 360

Received: 28 Feb 2024
Accepted: 08 Jul 2024
Published online: 12 Dec 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article