Our project delivers a machine learning innovation to revolutionize self-checkout systems in the retail sector by introducing YOLOv8 [1] on the Jetson Nano that proficiently identifies and tallies a wide range of produce. This system tackles the challenge of manual entries for unbarcoded items, a notorious bottleneck in retail workflows, and streamlines the checkout process for enhanced efficiency. By harnessing an extensive dataset of high-quality imagery and applying advanced optimizations like TensorRT and INT8 quantization, the model significantly curtails latency, particularly for larger models, crucial for seamless real-time application in bustling store settings. The implementation exhibits remarkable accuracy and precision in classifying produce, optimizing both latency and memory usage, thus pushing the envelope in retail automation. The successful quantitative outcomes point towards a robust, adaptive system that not only improves current operations but also sets the stage for future advancements in on-device machine learning applications within the retail industry.
Our self-checkout system features the NVIDIA Jetson Nano and the Logitech StreamCam, as illustrated in Figure 1. The Jetson Nano, with its 5V/4A power input, excels in processing efficiency, facilitating rapid and precise item detection. Its prowess in handling complex image tasks seamlessly integrates with the StreamCam, which offers Full HD clarity via USB-C, mounted on an Acetaken StreamCam Stand for adjustable scanning angles. Products are placed on a flat white styrofoam board for uniform background during scanning. The system’s interface is displayed on a 14-inch ARZOPA Portable Monitor Kickstand, ensuring a crisp FHD 1080P output with minimal power consumption, under 15W. This setup, combined with model quantization, achieves a high-accuracy, low-latency customer experience. Collectively, these components underscore our commitment to delivering a precise and user-friendly self-checkout solution.
For more information, please contact the authors:
- sjagades@andrew.cmu.edu
- oalama@andrew.cmu.edu
- aalajlan@andrew.cmu.edu