Sparse Semantic Encoding for Reduced Data Load in Vision-Position Aided mmWave Beam Prediction

To improve beam prediction in complex environments, millimeter-wave (mmWave) communications systems are increasingly using multi-modal sensing, leveraging environmental data from different sources such as visual and position sensors to enhance communications. Although this fusion enables a comprehensive view of the environment, transmitting raw data from a sensing node to the base station (BS) via wireless links incurs significant data overhead, presenting a bottleneck for real-time deployment. This paper presents an efficient machine learning (ML) framework that extracts sparse semantic features from RGB images to reduce data transmission while ensuring robust beam prediction. A custom loss function enforces sparsity in the transmitted features, and the combined sparse and position data are employed for beam prediction via an ML model. Experimental results show that our framework achieves comparable accuracy to more complex models but with a fraction of the computational and transmission costs.