This is the Retail Automation demo on AM62A Processor, which was used at Embedded World 2023 to showcase this new SoC.
| DEVICE | Supported |
|---|---|
| AM62A | ✔️ |
This demo is supported on EdgeAI SDK Version 9.0
Run run_demo.sh to run this demo.
This demo runs on the AM62A Starter-Kit EVM. It requires a 1920x1080 HDMI display and a camera. The camera can be an IMX219 module with CSI connection or a 720p/1080p USB camera. Copy the files within this directory (and subdirectories) or clone the repo to the device - it is not important where within the devices file system.
The run_demo.sh script assumes an IMX219 is attached to the CSI port on the device and that the model to use for inference is at food-detection-model-mobv2ssd. If the trained food recognition neural network is not already present at food-detection-model-mobv2ssd, it will be downloaded - this requires network connection to the processor. The base python file is retail_vision_app.py. The -h tag will print the full readout for command-line options. Additional command line tags can be added to the run_demo.sh call, which will be passed on to the python command. By default, the demo assumes an IMX219 CSI2 camera, but this can be changed with command line options
This application was developed using edgeai-gst-apps as a starting point for a smart retail demo. Foods are recognized with deep learning to fill a customer's order quickly and autonomously. Images are captured, preprocessed, inferenced using a custom-trained neural network, and are displayed to the screen with text emulating a receipt readout.
This application was developed for the AM62A, a quad-core microprocessor from Texas Instruments with 2 TOPS of deep learning acceleration, integrated ISP, and video codec. The food-recognition neural network is compiled for this deep learning accelerator. If using a different TI Edge AI Processor, the neural network will need to be recompiled. See TI's Edge AI repo and edgeai-tidl-tools to learn more about this process.
This demo is written in Python3 and leverages the ONNX runtime with TIDL Execution Provider to utilize the deep learning engine. It uses gstreamer with TI's custom plugins for accessing hardware accelerators within the SoC as well as OpenCV for additional image processing. The IMX219 is supported as a raw camera input sensor, as are 1080p and 720p USB cameras (logitech c920 and c270, respectively). Output is displayed to an HDMI screen, and 1920x1080 resolution is assumed.
See the FAQ
See the guide on reproducing the demo to find all compoents used in the making and displaying of this demo, including the food items the model was trained to recognize as well as mounting hardware. This includes basic guidance for compiling the model to run on different SDK versions, since the model may not be forwards/backwards compatible.
To learn more about how this demo was made, please see HOW_ITS_MADE.md