چكيده به لاتين
This thesis provides a detailed search into the real world of 3D point cloud registration, exploring three distinct methods for achieving precise alignment. The methods under scrutiny include the traditional conventional registration approach, a neural network aided by Geo-transformers, and the cutting-edge Deep Learning 3D Transformers method. Our investigation encompasses a comprehensive evaluation of these methods based on factors like precision, speed, noise resistance, and convergence characteristics. Traditional registration methods, rooted in mathematical optimization, offer elegance and accuracy in well-conditioned scenarios. However, traditional methods may grapple with challenges posed by noise, varying densities, and complex transformations. Neural network-based methods exhibit remarkable adaptability and noise resistance, learned through exposure to diverse point-cloud pairs. Nevertheless, they demand substantial computational resources and time for training and inference. The revolutionary Deep Learning 3D Transformers approach combines the power of transformers with point cloud registration. It effectively handles complex alignments and noise-ridden data by considering global contexts. Although this approach shines in scenarios where traditional methods fall short, it requires refinement for quicker convergence and improved computational efficiency. Comparing the methods highlights their respective strengths and limitations, fostering a deeper understanding of their applicability. The findings resonate with prior research, reinforcing the efficacy of deep learning in navigating intricate point cloud data. Recommendations for the development of the Deep Learning 3D Transformers method include architectural enhancements, computational efficiency improvements, and hybrid strategies. Moreover, this Thesis sheds light on the diverse landscape of 3D point cloud registration methods. Each method contributes to the field’s evolution, offering solutions ranging from well-established mathematical foundations to groundbreaking deep learning innovation. As challenges evolve, the synergy between tradition and innovation will drive the advancements shaping the future of point cloud registration.
