Embedded Oriented Object Detection on MaixCAM Pro for Waste-Sorting Perception

For sorting mechanisms that operate on a planar platform, the perception module may need to estimate object category, image-plane position, object extent and principal-axis orientation within the computing budget of a low-power device. This paper reports an oriented object detection workflow for a MaixCAM Pro (Shenzhen Sipeed Technology Co., Ltd., Shenzhen, China) waste-sorting prototype. The sensing output is represented by oriented bounding boxes (OBBs), which provide two-dimensional position, rectangular extent and principal-axis orientation for downstream sorting control. To adapt the detector to a small self-built dataset and INT8 deployment, class- and aspect-ratio-aware re-sampling (CAR-RS) repeats selected training images, while class- and aspect-ratio-aware calibration (CAR-Calib) selects a stratified calibration subset for post-training quantization. The YOLO11-OBB model is exported as a static ONNX graph, adapted by exposing the OBB head output tensors, compiled with TPU-MLIR and executed through native MaixPy OBB inference. On the original PC validation split, the baseline and CAR-RS models both obtain 0.995 mAP50, with mAP50–95 values of 0.935 and 0.932, respectively. CAR-RS reduces the angle mean absolute error from 3.403° to 3.201° for objects with aspect ratio greater than or equal to 1.3. On an additional 80-image cross-background test set that was not used for training or calibration, CAR-RS improves precision from 0.839 to 0.907, recall from 0.882 to 0.893 and mAP50 from 0.916 to 0.927, while mAP50–95 changes from 0.797 to 0.789. On a 100-image MaixCAM Pro validation set, the CAR-RS + CAR-Calib INT8 model achieves 0.961 precision, 0.994 recall, 0.977 F1-score and 0.980 mAP50. The deployed INT8 model occupies 11.85 MB and runs at 8.40 FPS. The results support the technical feasibility of executing an OBB detector on MaixCAM Pro, while also showing that independent-scene validation remains necessary for broader deployment.