Comparison of DSO and ORBSLAM3 in Indoor Environments Using Monocular Visual SLAM

Abstract

This thesis presents a comparative study of ORB-SLAM3 and DSO in indoor environments. We first validate the local implementation on public datasets, then analyze self-recorded sequences from four perspectives: trajectory behavior, summary statistics, average per-frame runtime, and mapping visualization. The results show that ORB-SLAM3 is stronger in pose estimation and real time efficiency, while DSO provides more informative semi-dense map appearance. Both methods degrade under low light and strong highlights, but with different failure modes, indicating clear complementarity. The study also notes key limitations, including the absence of ground truth for custom data, limited run repetitions, and acquisition choices that increase loop-closure difficulty. Future work will expand scene coverage and sample size, improve the evaluation protocol, and explore a hybrid framework that combines the strengths of both approaches. Keywords: semantic communication, vehicle detection, YOLOv12, CNN encoder, supernet, AWGN channel, compression ratio.