Abstract
Employing point cloud and 3-D data has recently proliferated with the advent of emerging Internet of Things (IoT) applications, e.g., Autonomous Vehicles (AV), and Augmented Reality/Virtual Reality (AR/VR), with their resource-limited IoT devices equipped with Light Detection and Ranging (LiDAR) sensors. While offloading the point cloud computations could improve energy-efficiency, the imposed bandwidth and energy dissipation due to transmitting a significant volume of 3-D data is still challenging. This issue gets more complicated due to the instability caused by the noisy channel conditions, which leads to the cliff effect and reduced object detection accuracy. Accordingly, this article introduces NORRIS; a semantic-based point cloud data transmission technique that mitigates transmitted data while enhancing resilience against channel noise. NORRIS utilizes joint source-channel coding in combination with transfer learning. It emphasizes efficiency and performance, providing a lightweight model that sustains excellent performance across different channel conditions, making it ideal for resource-constraint IoT devices. While NORRIS provides high detection accuracy, it significantly reduces the inference time and model complexity. The effectiveness of NORRIS is evaluated and compared with state-of-the-art through a comprehensive set of experiments. Results demonstrate that NORRIS successfully overcomes the cliff effect while guaranteeing a higher performance in lower Signal-to-Noise Ratio (SNR) conditions. It also reduces the transmitted data volume by up to $10\times $ , while accelerating meaning extraction by $5\times $ . It also enhances accuracy by approximately 10% in various SNR scenarios. Furthermore, NORRIS maintains stable classification performance under noise, with only a 5% accuracy drop, in contrast to the 60% decrease observed in existing approaches.