MVSOP: A new framework for integrating MVS into SLAM based on ORB and PF

Currently, the sparse map generated by simultaneous localization and mapping (SLAM) based on features is insufficient for practical use. To acquire a dense map, the model of dense reconstruction is usually appended to the SLAM system. One of the issues is that the dense reconstruction is fully independent of the SLAM system. In other words, the dense reconstruction hardly utilizes any information from the SLAM. To sufficiently integrate the dense reconstruction with traditional visual SLAM frameworks, in this paper, a new framework for integrating dense reconstruction into SLAM based on ORB and PF is proposed. It combines the ORB features that have been utilized in the visual odometry of SLAM with multi-view stereo (MVS) for dense map reconstruction. To incorporate ORB features, a new particle filter model is designed. The model includes an observation model based on ORB features to rectify the estimation errors by the MVS method. In addition, A new asynchronous fusion algorithm is proposed to update the proposed model. Finally, the proposed framework is evaluated on a public dataset that includes several subsets, which demonstrates its rationality and effiency.