Semantically Assisted Lightweight SLAM Based on Embedded Devices

The rise of edge computing has increased interest in deploying SLAM (Simultaneous Localization and Mapping) technology on embedded systems. However, traditional SLAM algorithms face challenges in achieving both real-time performance and accuracy on resource-constrained platforms. This paper proposes a lightweight, semantically-assisted SLAM framework designed to address these challenges. The framework enhances system performance by optimizing semantic segmentation, dynamically clustering feature points, and introducing an adaptive weighting mechanism. We introduce key optimizations, including the pruning and quantization of the SalsaNext semantic segmentation model to reduce computational complexity. Additionally, a Gaussian Mixture Model (GMM) is employed to filter and cluster feature points. A dynamic weighting strategy, based on historical frame matching and semantic context, further improves accuracy in dynamic scenarios. Experimental results on the KITTI dataset show that the proposed framework achieves competitive localization accuracy compared to state-of-the-art SLAM systems, with significant improvements in computational efficiency and real-time performance. This work presents a practical approach for developing efficient SLAM solutions on embedded devices.