MoV-SLAM: Using Motion Vectors for Real-Time Single-CPU Visual SLAM

The 6 degrees-of-freedom movement of micro-aerial vehicles (MAVs) combined with their lightweight, low-power computing makes accurate, real-time (RT) visual simultaneous localization and mapping (vSLAM) for MAVs considerably challenging. Existing classical and learning-based SLAM approaches require high computing capabilities to achieve RT performance, which prevents long-range deployment on lightweight, lowpowered, and low-resource devices. We present MoV-SLAM, an approach that obtains long-range robust correspondence tracks for vSLAM using motion vectors that link macroblocks decoded from H.264 video coders. We contribute efficient algorithms to detect and match distinctive blocks for fast, accurate, and reliable vSLAM on low-resource, single-processor systems in MAVs. Our approach outperforms the existing state-of-the-art (SOTA) for single-processor vSLAM reconstruction when evaluated on multiple off-the-shelf datasets. Project page: https://github.com/Terascale-All-sensing-Research-Studio/MOV-SLAM