Real-time monocular 3D reconstruction of scenarios using artificial intelligence techniques

Resumen

This research presents a comprehensive study on monocular 3D reconstruction of environments using only RGB images as input acquired through a monocular sensor. The objectives were to develop a suitable taxonomy, review seminal algorithms, compare open-source methods, and develop a novel 3D reconstruction system using the principal classic techniques combined with artificial intelligence to improve the overall system performance. An exhaustive literature review led to a proposed taxonomy with three classifications: direct vs indirect, dense vs sparse, and classic vs machine learning. This resulted in 10 categories used to classify 42 notable monocular SLAM, SFM, and VO systems based on 11 identified criteria. Subsequently, through rigorous benchmarking, ten prominent open-source algorithms were implemented across the taxonomy to discern each method's advantages and limitations. The TUM-Mono dataset, considered the most complete benchmark comprising 50 outdoor and indoor sequences, was used for evaluation. Statistical analysis revealed that sparse-direct methods significantly outperformed others, with DSO excelling. In addition, it was evidenced that integrating machine learning modules into the SLAM pipeline significantly contributes to the system performance and the final reconstruction quality. Consequently, DSO was selected for enhancement by integrating the stateof- the-art single image depth estimation NeW-CRFs CNN module. This module introduced depth prior knowledge to refine DSO's depth initialization and tracking. Using the TUM-Mono dataset, the new DeepDSO method was benchmarked against DSO and CNN-DSO. DeepDSO surpassed the others across various metrics, including translation error, rotation error, scale error, alignment error, and RMSE. Statistical tests confirmed DeepDSO's superiority, achieving an impressive RMSE of 0.0624, which corresponds to an error reduction close to 13.35% with respect to the original DSO system. DeepDSO pushes monocular VO boundaries by strategically integrating machine learning-based depth estimation. In addition, the taxonomy and comparative analysis provide guidelines for appropriate algorithm selection and implementation. This study validates the benefits of implementing artificial intelligence within SLAM, VO and SFM systems and lays the groundwork for continued depth initialization and point-tracking optimisations.