QoE-Aware Latency Optimization in Semantic Transmission Empowered Edge Assisted AR

Augmented reality (AR) enhances human perception by overlaying digital information in the real environment and can be applied to various fields such as education, healthcare, and industry. To relieve the computing pressure of resourcelimited AR devices, edge assisted AR has attracted significant attention in recent years. It enables AR devices to offload the computation-intensive environment analysis tasks to the edge server and to download the point-based 3D virtual objects for rendering. Although this method can effectively reduce the processing latency, it may introduce a long transmission latency. To address this issue, this work proposes a semantic communication empowered edge assisted AR to reduce the quality of experience (QoE)-aware latency. To this end, the AR devices first utilize an image encoder to extract semantic features of source images. Due to the importance difference of extracted features and limited communication resources, the feature selection and bandwidth allocation are explored to improve transmission efficiency. In addition, when the edge server completes the generation of pointbased 3D virtual objects, it employs a transformer-based encoder to mask a proportion of points and learn a high-level representation of unmasked points. In this way, low-complexity semantic information of 3D objects is sent to the AR devices, which can reduce the transmission overhead of 3D objects. Moreover, the mask ratio of the point encoder is further optimized to balance the reconstructed quality and transmission latency of 3D objects. Therefore, this work aims to minimize the overall QoE-aware latency of edge assisted AR by jointly optimizing the semantic feature selection, mask ratio, and bandwidth allocation. A convexbased improved Harris Hawks optimization algorithm is designed to solve the formulated problem. Experimental results demonstrate that our semantic communication empowered edge assisted AR can effectively reduce the overall latency and outperform the other baselines