This paper presents a lightweight MCU–EdgeTPU platform—a microcontroller unit (MCU) paired with an Edge Tensor Processing Unit (EdgeTPU) accelerator—for onboard drone-perception experiments, extended from an open-source baseline originally limited to Quarter Video Graphics Array (QVGA) single-camera operation. Rather than treating hardware, runtime, model, and data as separate problems, they are developed as parts of the same continuous perception pipeline. The platform extends the hardware baseline toward dual 5 Mpx sensing, onboard inertial measurement unit (IMU) support, real-time embedded inference, and a high-level MicroPython control layer. In parallel, lightweight You Only Look Once (YOLO) detectors are trained and selected on a synthetic aerial-person dataset generated under the visual conditions expected by the drone camera, including target resolution, viewpoint, object scale, weather, lighting, and time-of-day variation. The resulting workflow starts from both ends: the detector must be small and quantization-stable enough for the EdgeTPU path, while the dataset must match the images that the onboard sensor is expected to observe. To evaluate the system, the full path from camera capture and image conversion to TPU transfer, model execution, and post-inference processing is analyzed. In the tested setup, the optimized single-camera pipeline runs stably with no timeouts or inference failures at about 26 detections per second with standard RGB input; because each EdgeTPU invocation is bounded by the USB transfer of the input image, feeding the camera’s native YUV420 format instead halves that transfer and raises throughput to about 40 detections per second at the same accuracy, while the selected 8-bit-integer (INT8) person detector preserves most of its 32-bit floating-point (FP32) accuracy. Detections are exposed to drone-control workflows (MAVLink/PX4 and Crazyflie) through the scriptable layer as an integration interface rather than a validated autonomy stack. The central contribution is therefore a co-designed embedded perception pipeline in which the board, runtime, detector, dataset, and even the camera pixel format are aligned around the same operating conditions.
Nedelcu et al. (Mon,) studied this question.