Virtual View Generation Based on 3D-Dense-Attentive GAN Networks.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Binocular vision systems are crucial for intelligent vehicles but are sensitive to missing visual data.
• Existing systems lack robustness when encountering signal loss, impacting vehicle perception.
• Generative adversarial networks (GAN) offer potential for creating realistic synthetic data.

Purpose of the Study:

• To propose a novel virtual view-generation algorithm to enhance the robustness of binocular vision systems.
• To address the challenge of missing vision signals in intelligent vehicle perception.
• To improve the reliability and performance of autonomous driving systems.

Main Methods:

• A generative adversarial network (GAN) framework comprising generative and discriminator networks.
• A generative network utilizing 3D convolutional neural networks (3D-CNN) and attention mechanisms for time-series feature extraction.
• A discriminator network employing dense block structures to prevent gradient vanishing and incorporating image edge, depth map, and optical flow constraints.

Main Results:

• The proposed algorithm successfully generates virtual views to replace missing vision signals.
• Quantitative evaluations on KITTI and Cityscapes datasets show superior performance compared to conventional methods.
• The generated virtual views effectively enhance the robustness of binocular vision systems.

Conclusions:

• The GAN-based virtual view-generation algorithm significantly improves the robustness of binocular vision systems.
• This approach offers a viable solution for handling missing vision signals in intelligent vehicles.
• The method demonstrates potential for advancing autonomous driving perception capabilities.