A Deep Learning Framework for Accurate Vehicle Yaw Angle Estimation from a Monocular Camera Based on Part Arrangement

Area of Science:

• Computer Vision
• Deep Learning
• Robotics

Background:

• Accurate object pose estimation is crucial for state assessment and movement prediction.
• Current methods often rely on 3D model matching or direct pose regression, potentially overlooking the significance of object part configurations.
• Early computer vision research highlighted the strong correlation between object parts and overall pose.

Purpose of the Study:

• To propose a novel, lightweight deep learning framework named YAEN (yaw angle estimation network) for accurate object yaw angle prediction.
• To leverage the arrangement of object parts for improved yaw angle estimation from monocular camera input.
• To develop a computationally efficient model suitable for real-time applications with low hardware requirements.

Main Methods:

• Developed an encoding-decoding deep learning structure (YAEN) for vehicle yaw angle prediction.
• Utilized a part-encoding network to extract vehicle part arrangement information.
• Employed a yaw angle decoding network to estimate the yaw angle from the extracted part information.
• Incorporated asymmetric convolution and a signed sum of squared errors (SSE) loss function to enhance model performance.
• Created a real-world dataset using vehicles with high-precision positioning devices for model validation.

Main Results:

• The YAEN model demonstrated a lightweight design with low hardware requirements, achieving a detection speed of 97 FPS on a 2070s graphics card.
• Experimental results showed satisfactory prediction performance in scenarios without vehicle occlusion.
• Achieved an average prediction error of less than 3.1 degrees.
• Attained an accuracy of 96.45% for prediction errors within 10 degrees in real driving scenarios.

Conclusions:

• The proposed YAEN framework effectively predicts vehicle yaw angles using part arrangement information from monocular cameras.
• The model's lightweight architecture and high accuracy make it suitable for real-time applications in autonomous driving and robotics.
• The study validates the importance of considering object part configurations in pose estimation tasks.