Metric-scale monocular SLAM via lightweight depth-integrated and PSO-calibrated ORB-SLAM3

Monocular Simultaneous Localization and Mapping (SLAM) has gained increasing attention for low-cost and lightweight deployment in robotics and autonomous systems. However, conventional monocular SLAM pipelines often lack metric scale and rely on depth networks that are trained on a single dataset, limiting their generalization and applicability to resource-constrained platforms. In this study, we present an extension of ORB-SLAM3 with a depth-agnostic integration framework initially tested with several modern monocular depth predictors, including MiDaS, Depth Anything V2, RT-MonoDepth, and Lite-Mono. Among these, Depth Anything V2 showed the most robust performance on our collected dataset, benefiting from its training on diverse data and state-of-the-art design. To address the mismatch between relative depth outputs and metric requirements, we propose a four-parameter disparity-to-depth mapping optimized using Particle Swarm Optimization (PSO), which we apply to Depth Anything V2 in our final configuration. The PSO-based calibration is performed once per deployment environment using a single representative sequence; the resulting parameters are then fixed and applied to all remaining sequences without re-optimization. Experiments were conducted on the KITTI benchmark, a publicly available IUST car dataset, and newly collected phone-mounted driving sequences. Using Depth Anything V2 as the backbone, our system achieved real-time performance on a modest GTX 1050 laptop, demonstrating feasibility for Jetson-class embedded hardware. In our method, the trajectory scale is fixed to 1.0 for all sequences, whereas original ORB-SLAM3 and LFFS require per-sequence scale estimation before error computation. On the KITTI odometry benchmark, the proposed PSO-calibrated framework significantly outperforms baselines. For instance, on sequence 00, ATE is reduced from 128.0740 m (ORB-SLAM3) and 69.1908 m (LFFS) to 4.5642 m (Ours). Similar improvements are observed across other sequences: sequence 02 (155.5040 m → 12.4065 m), sequence 05 (91.8820 m → 7.0894 m), sequence 07 (14.6667 m → 2.6175 m), and sequence 10 (30.2473 m → 1.2800 m). On the IUST driving dataset, ATE drops from 48.5212 m (ORB-SLAM3) to 9.4257 m. For the newly collected smartphone-mounted dataset under challenging real-world conditions (motion blur, dynamic traffic, suburban driving), the system maintains stable tracking with ATE values of 9.2483 m (sequence 00) and 13.5612 m (sequence 01) against GPS reference, while baselines fail to produce reliable trajectories. These results highlight improved scale stability and accuracy compared to fixed mappings, with latency compatible with real-world deployment.