Work-in-Progress: Smart data reduction in SLAM methods for embedded systems

Abstract

Visual-inertial simultaneous localization and mapping methods (SLAM) process and store large amounts of data based on image sequences to estimate accurate and robust real-time trajectories. Real-time performances, memory management and low power consumption are critical for embedded SLAM with restrictive hardware resources. We aim at reducing the amount of injected input data in SLAM algorithms and, thereby, the memory footprint while providing improved real-time performances. Two decimation approaches are used, constant filtering and adaptive filtering. The first one decimates input images to reduce frame rate (from 20 to 10, 7, 5 and 2 fps). The latter one uses inertial measurements to reduce the frame rate when no significant motion is detected. Applied to SLAM methods, it produces more accurate trajectories than the constant filtering approach, while further reducing the amount of injected data up to 85%. It also impacts the resource utilization by reducing up to an average of 91% the peak of memory consumption.