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

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Sorting mechanisms on planar platforms require efficient object perception, including category, position, extent, and orientation estimation within low-power constraints.
• Existing object detection methods often struggle with real-time performance and resource limitations on embedded systems.

Purpose of the Study:

• To develop and evaluate an oriented object detection workflow for a MaixCAM Pro waste-sorting prototype.
• To adapt an object detector for small datasets and INT8 deployment on resource-constrained hardware.
• To enable downstream sorting control using oriented bounding boxes (OBBs).

Main Methods:

• Implemented a YOLO11-OBB model exported as a static ONNX graph with exposed OBB head output tensors.
• Utilized class- and aspect-ratio-aware re-sampling (CAR-RS) for adapting the detector to a small dataset.
• Employed class- and aspect-ratio-aware calibration (CAR-Calib) for post-training quantization to INT8.
• Compiled the model with TPU-MLIR and executed inference using native MaixPy OBB.

Main Results:

• The CAR-RS model achieved competitive performance on the PC validation split, with mAP50 of 0.995 and mAP50-95 of 0.932.
• CAR-RS significantly reduced the angle mean absolute error for objects with aspect ratios >= 1.3, from 3.403° to 3.201°.
• On a cross-background test set, CAR-RS improved precision to 0.907 and recall to 0.893.
• The final INT8 model on the MaixCAM Pro achieved 0.977 F1-score and 0.980 mAP50, running at 8.40 FPS and occupying 11.85 MB.

Conclusions:

• The study confirms the technical feasibility of deploying an oriented object detection model on the MaixCAM Pro for waste sorting.
• The proposed CAR-RS and CAR-Calib methods effectively adapt detectors to small datasets and INT8 deployment.
• Independent-scene validation is crucial for ensuring the broader applicability and reliability of deployed models.