<![CDATA[Industrial automation systems require automatic, real-time object counting on conveyors. Conventional sensors have limitations in handling variations in object shapes and conditions. A deep learning-based computer vision approach can be a solution, but its implementation on embedded devices such as the Raspberry Pi 5 faces computational limitations that result in slow inference speeds. This research implements the YOLOv8s model, accelerated using the Neural Processing Unit (NPU)-based Hailo-8L on the Raspberry Pi 5, for real-time object counting. A dataset of 1,343 bottle cap images was annotated using Roboflow and divided into 70% for training, 20% for validation, and 10% for testing, with a resolution of 640 x 640 pixels. The model was trained using Google Colab, exported to ONNX format, and compiled into Hailo Executable Format (.hef) using the Hailo Dataflow Compiler. Testing was conducted at conveyor speeds of 1.6–10.0 m/min with and without the Hailo-8L. Test results showed an increase in FPS from an average of 2.2 FPS to 28.9 FPS—approximately a 13.1-fold increase. In controlled counting tests, the system with Hailo-8L achieved up to 100% accuracy, while the system without Hailo-8L experienced a decline in counting accuracy at higher conveyor speeds. Validation in an industrial environment demonstrated a counting success rate of 97%. These results indicate that the Hailo-8L’s acceleration capability can reduce the YOLOv8s inference computational bottleneck and improve the real-time performance of the counting system.]]>
