A siamese network-based approach for vehicle pose estimation.

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Accurate vehicle pose estimation is crucial for autonomous systems.
• Monocular camera-based methods face challenges in depth perception and scale ambiguity.

Purpose of the Study:

• To propose a novel deep learning model, FPN PoseEstimateNet, for accurate vehicle pose estimation using monocular cameras.
• To enhance feature scale handling and enable independent training of network components.

Main Methods:

• A Siamese network-based feature extractor with a Feature Pyramid Network (FPN) was employed.
• A correlation matrix was generated for feature matching between images.
• A pose calculation network predicted vehicle pose changes based on correlation and standard matrices.

Main Results:

• The FPN PoseEstimateNet achieved an angle error within 8.26° and a maximum translation error within 31.55 m.
• The network operates at 6 FPS with a parameter size of 101.6 million.
• The feature extractor was trained independently using time intervals as labels.

Conclusions:

• FPN PoseEstimateNet offers an effective deep learning solution for monocular vehicle pose estimation.
• The method demonstrates robust performance across different sequences.
• The architecture facilitates efficient feature extraction and pose prediction.