Comparison of the YOLO11n model on embedded platforms Orange Pi5 and Raspberry Pi4B

Abstract

Context and relevance. Modern trends in technology development are moving towards the transfer of computing from cloud platforms to embedded devices. In this regard, there is a growing need to determine the optimal stack of hardware and software technologies for solving computer vision problems. Goal. Determine the optimal hardware configuration for real-time computer vision tasks based on the criteria of bandwidth, latency, accuracy, and ease of integration. Methods and materials. The paper provides a direct experimental comparison of two embedded platforms for executing the YOLO11n object detection model: a single-board Orange Pi 5 computer with a built-in Rockchip RK3588S neuroprocessor (6 TOPS, INT8) and a Raspberry Pi 4B bundle with an external Google Coral Edge TPU USB accelerator (4 TOPS, INT8). The YOLO11n model was converted from a single PyTorch source file to the target formats RKNN and TensorFlow Lite (full integer quantization) using the tool chains recommended by the manufacturers. The measurements were performed under identical pretreatment conditions. Results. The Orange Pi 5 achieved a steady frame rate of 43-54 FPS (640×640 resolution) with an end-to-end delay of 19-23 ms; accuracy mAP@0.5:0.95 with INT8 quantization, it remains at ~39%. On a Raspberry Pi 4B with Edge TPU at a resolution of 448×448, the delay was 121.5 ms (~8.2 FPS), while 45% of operations are performed on the CPU, which limits scaling. Conclusions. The Orange Pi's gain in speed reaches 6-7 times at a higher resolution, the NPU demonstrates a significant load margin, and the single-board implementation eliminates the overhead of the external interface. The totality of the results allows us to recommend Orange Pi 5 as the preferred platform for building productive edge solutions with real-time object detection.