EVS-SLAM: a lightweight visual SLAM system for autonomous robots in dynamic environments

Human-Intelligent Systems Integration · Springer Nature · 2026-10-26 · DOI: 10.1007/s10514-026-10273-4

Аннотация

Dynamic objects challenge Visual Simultaneous Localization and Mapping (VSLAM) by introducing observations that violate the static-scene assumption used for pose estimation and mapping. This paper presents EVS-SLAM, a modular RGB-D/stereo VSLAM system designed for deployment on embedded companion computers. EVS-SLAM integrates DeepLabV3-ResNet50-EO, a deployment-adapted implementation of DeepLabV3 with a ResNet50 backbone, on the NVIDIA Jetson AGX Xavier. Semantic masks classify ORB observations as static or potentially dynamic, allowing the latter to be excluded before feature matching, camera-pose estimation, and map construction. Evaluation on the KITTI and TUM benchmarks shows competitive trajectory accuracy relative to representative dynamic-scene VSLAM systems. In a controlled same-platform ablation, jointly enabling semantic segmentation and dynamic-feature filtering reduced the mean ATE RMSE by 92.84–$$97.42\%$$ on the TUM RGB-D walking sequences. On the selected KITTI stereo sequences, APE RMSE changes ranged from a $$5.72\%$$ improvement to a $$1.93\%$$ degradation and were comparable to run-to-run variability, demonstrating that the benefit is scene-dependent. Concurrent CPU–GPU execution yielded effective processing throughputs of 13.76 FPS on KITTI and 19.20 FPS on TUM, while dynamic-feature filtering itself accounted for less than $$1\%$$ of total latency. Qualitative indoor, urban, and parking-lot deployments further demonstrate onboard operation in real-world environments.

Читать у издателя →

Похожие статьи