This application, included within the sample projects, enables hardware-accelerated object detection using
the C7x/MMA cores on the AM67A SoC. The camera feed is captured via the GStreamer framework; inference is
performed using a TensorFlow Lite model, and the results are streamed in real-time to a browser via an MJPEG server.The application uses the TIDL (TI Deep Learning) delegate library to run model inference on the C7x DSP.
If the TIDL delegate is not available, it automatically falls back to running on the CPU.The source code is available in T3 Gemstone’s GitHub
examples repository.
This demo showcases an audio-visual system that processes both audio and video data. It detects voice commands
to control the camera’s focus area and uses deep learning to recognize and display faces on the screen.The source code is available on Texas Instruments’ GitHub
edgeai-demo-audio-visual repository.
Barcodes play a critical role in areas such as inventory management, asset tracking, ticketing, and information
sharing. One-dimensional (1-D) and two-dimensional (2-D) barcodes condense information into a visual encoding
format. While laser-based scanners are sufficient for 1-D barcodes, cameras are required for 2-D barcodes
(e.g., QR codes). Camera-based systems are often called “barcode imagers.” In barcode imagers, the most
computationally intensive task is not decoding the barcode but rather locating it, and deep learning
techniques are highly effective at this stage.This demo runs a specially trained YOLOX-nano neural network to detect 1-D and 2-D barcodes. The detected
barcode regions are cropped and converted to grayscale for decoding using the open-source zbar library.
The decoded barcode is displayed alongside the bounding box obtained from object detection.The source code is available on Texas Instruments’ GitHub
edgeai-gst-apps-barcode-reader repository.
Self-checkout systems in retail and grocery stores have significantly improved the customer experience.
Simple and user-friendly kiosks allow customers to scan their products and make low-contact payments. This
field is increasingly moving towards more automated systems, where customers can scan multiple items
simultaneously without needing to search for barcodes or position items in a specific way. Such systems
offer additional advantages like speeding up the process, reducing contact surfaces, and increasing
overall efficiency.This demo showcases an automated retail checkout scanner that uses a deep learning model to detect 12 different
types of food items (banana, apple, chip bag, soda can, etc.).The source code is available on Texas Instruments’ GitHub
edgeai-gst-apps-retail-checkout repository.
Image-based people tracking can be applied in many fields such as retail, building automation, security, and
safety. This demo uses the YOLOX-S-Lite machine learning model to detect individuals in a video stream. The
model’s output is used to track the movements of people in the scene, leveraging the open-source Norfair library.The demo provides live tracking of individuals in the scene and displays timers showing the duration each
person has spent in their current location. It also includes a control panel showing statistics such as total
visitor count, current occupancy, and the distribution of time individuals spend in the scene.
Additionally, a heatmap highlights frequently visited areas. This feature provides valuable insights
for understanding human behavior. For example, in retail stores, it can help optimize shelf layout to
improve customer experience.The source code is available on Texas Instruments’ GitHub
edgeai-gst-apps-people-tracking repository.
