Implementation and testing of a 5G-based intelligent IoT system for image recognition

Integration of Artificial Intelligence into the Internet of Things is fundamentally changing how edge devices process and transmit data. In traditional computer vision networks, cameras act as simple data pipelines, transmitting video streams to cloud servers. Although 5G networks offer high bandwidth, this data-oriented transmission wastes a lot of infrastructure energy, saturates the radio uplink, and causes delays. This thesis investigates a Goal-Oriented (GO) communication paradigm, where intelligent edge devices process information locally and initiate transmissions only when task-relevant data is detected. To evaluate this approach, a distributed computer vision testbed was designed using Raspberry Pi microcomputers equipped with Quectel 5G modems, Ultralytics YOLO object detection models, and a private 5G network powered by an OpenAirInterface. Three strategies were tested: Device-Only, Cloud-Only, and Hybrid (GO). To correlate software logic with hardware limits, a telemetry system utilizing PZEM-004T analyzers was implemented, synchronizing inference latency with the physical power consumption of the edge nodes, the 5G network, and the cloud server. Outdoor signal evaluation proved the necessity of the GO paradigm: as physical distance pushed the Reference Signal Received Power below -115 dBm, the 5G uplink collapsed, suffering ~90% packet loss and rendering Cloud-Only offloading ineffective. Experimental results demonstrate that deploying an intermediate YOLO model at the edge successfully acts as a semantic gatekeeper. By dropping uninformative frames locally, the GO architecture reduced the baseline End-to-End latency to 200 ms and successfully protected the cloud server from idle power spikes. However, executing heavier filter models at the edge imposes a continuous 8-10 Watt thermal penalty on the local node, while the 5G base station remains the dominant infrastructure bottleneck, drawing a constant 120 Watts to maintain the active radio link.