Multi-modal feature extraction and depth prediction algorithm in intelligent visual analysis

View at publisher Summarise with VIRAT
Authors: Wenshen Xu
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In intelligent vision analysis, single-mode data are easily disturbed by illumination and occlusion, and the existing multi-mode fusion methods have high computational complexity and poor dynamic adaptability. This paper proposes an adaptive multi-mode feature fusion and dynamic depth prediction framework. The framework constructs a dynamic gated fusion network, extracts hierarchical features through modal-specific encoders, and introduces a gating mechanism to dynamically adjust the weights of each mode to alleviate redundancy and conflict. The spatio-temporal attention fusion module is designed, and the long-distance spatio-temporal dependency is jointly modeled by non-local operation and Transformer encoder to enhance the global feature interaction. A dynamic graph reasoning network is proposed to structure the scene into a spatio-temporal dynamic graph, and the hidden space evolution reasoning is realized by accumulating neighbor information through graph volumes and using gated cycle units, and the motion smoothness constraint and multi-task learning loss are combined to improve the generalization ability. Experiments on KITTI Vision Benchmark and Cityscapes data sets show that the MSE and MAE of the proposed algorithm reach 63.8 and 5.34 respectively in the depth prediction task, which is about 15% lower than that of the existing optimal methods, and mIoU reaches 71.8% in the semantic segmentation prediction task, which verifies the effectiveness of each module and the advanced nature of the framework in complex dynamic scenes.

Fields of Study & Subjects

Advanced Vision and Imaging Advanced Neural Network Applications Autonomous Vehicle Technology and Safety

Publication details

Journal / source
IET conference proceedings.
Volume
2025
Issue
47
Pages
868–873
Publication date
1 March 2026
Language
EN