Selective Frame Processing for Accelerating Visual SLAM
Abstract
Embedded deployment of visual and visual-inertial Simultaneous Localization and Mapping (SLAM) is often bot-tlenecked by the intensive computational overhead of processing every frame through descriptor extraction and standard front-end tracking, even when consecutive frames provide limited new information. This inefficiency results in long runtime and elevated energy consumption on resource-constrained platforms. To address these issues, we propose a cascaded frame-screening framework for ORB-SLAM2 and ORB-SLAM3 that targets redundant front-end computations. The ORB-SLAM2 policy leverages an ultra-low-cost keypoint-distribution heatmap to detect low-information frames, allowing frames with high similarity to bypass both Lucas-Kanade (LK) optical flow and descriptor-based tracking in favor of lightweight pose prediction. Ambiguous cases are subsequently evaluated by an LK-based second-stage gate to determine whether remaining front-end computation can also be skipped. For ORB-SLAM3, we further extend the cascade with inertial motion cues: frames with minimal Inertial Measurement Unit (IMU) motion and substantial heatmap overlap are classified as non-keyframes, while uncertain cases proceed through the LK gate and, if necessary, full tracking. Both policies are governed by keyframe-distance and tracking-state constraints. We evaluate the proposed framework on standard European Robotics Challenge Micro Aerial Vehicle (EuRoC MAV) sequences and embedded platforms in terms of trajectory accuracy, frame-culling rate, runtime, and energy consumption. The ORB-SLAM2 variant culls 26.64% of frames on average, while the ORB-SLAM3 Mono-Inertial variant reduces average desktop runtime by 19.17% with slightly improved average ATE RMSE. Embedded measurements show consistent efficiency gains, with mean per-sequence energy reductions of 13.07% on the ARM-based SoC and 15.30% on the NVIDIA Jetson AGX Xavier.