Bin Occupancy Estimation in Warehouses With a TinyML Vision Sensor

Machine vision has improved logistics efficiency, but its use for estimating available space in traditional warehouses remains limited, even with Artificial Intelligence of Things and tiny machine learning enabling cost-effective, energy-efficient connected sensors. This letter presents a sensor solution with two optimized You Only Look Once (YOLO) models for bin and box detection, using the detected dimensions to estimate bin occupancy and transmit the results over a wireless network. Both models were 8-bit quantized and embedded for efficient on-device processing. The bin detection model achieved an average precision of 0.937, and the box model 0.808, indicating reliable detection of medium and large objects. With mean prediction latencies of 1.697 s for bins and 1.728 s for boxes, the solution enables up to 184 days of battery autonomy with 30-min detection intervals, supporting smarter and efficient warehouse management.