An efficient stereo rectification model for autonomous driving

Abstract

Stereo rectification is a crucial step in autonomous driving, ensuring accurate depth estimation for safe and reliable navigation. Traditional rectification methods such as Hartley’s, Bouguet’s and Fusiello’s approaches offer geometric solutions but often struggle with computational efficiency and robustness in real-world scenarios. This research project hence proposes a deep-learning-based stereo rectification model leveraging EfficientNet-B0 and LSTM-based temporal refinement, trained on KITTI 2015 and Argoverse datasets. The model integrates a custom loss function combining Frobenius norm, epipolar constraint loss and determinant regularisation loss to improve homography estimation.

Experimental results demonstrate that the proposed model outperforms traditional rectification techniques in both accuracy and computational efficiency. It achieve an average processing time per pair that is better than these traditional methods by almost two folds, while reducing vertical disparity errors. Against deep learning models such as Deep3D and ECNN, the proposed model balances speed and accuracy, making it suitable for real-time applications.

This project hence bridges the gap between traditional and deep-learning-based rectification, offering an efficient and scalable solution for real-time stereo vision in autonomous vehicles, contributing to safer and more reliable depth perception.