Object detection in vision systems with 25x lower latency
Find and localize specific objects, people in real-time at high framerate with AI-accelerated processors and industry-standard software
Application overview
Many vision-based systems need to detect objects in the camera's field of view, like vehicles, people, faces, faulty parts and scannable codes. Knowing how many objects and where they are important in applications across factory automation, home/building security, automotive safety, robotics and more.
Complex objects require deep-learning AI solutions for reliable, accurate detection. Expert-level knowledge can be replaced with example images of the objects to detect, simplifying the development process. AI-based object detection is also more robust to variations in the scene, like lighting, unexpected objects, and sensor noise.
Integrating AI at the edge improves privacy and security of the system, and provides fast responses that are not limited by network speed or reliability. Hardware acceleration, like the C7™ NPU in TDA4x and AM62A devices, and industry standard tools, like ONNX Runtime, enable object detection open source AI models to run on-chip in real-time.
Starting evaluation
Data collection
Object detection training data consists of images and annotations, where the annotations describe which objects exist where in the image. TI tooling uses the MS COCO format for datasets.
Images for object detection should match the environment the end-system will run in. Capturing data with similar lighting conditions, camera lens, and variations in the environment will lead to higher accuracy of the trained model.
Data quality assessment
Image quality is important for high accuracy in the trained output, and there should be some degree of variation in the data.The objects you will detect should be shown in different ways, like turning the object, having it closer or further from the camera, or intentionally occluding part of the object so it is not entirely shown. More variation in the training data will ensure your model can handle such variations in the real world, although this is not strictly required.
A few dozen images per object type is sufficient to build a reasonably accurate model.
Build and train your model
We provide an optimized set of models for fine-tuning on custom datasets through CCStudio™ Edge AI Studio or with an analogous command-line tool on the official TI GitHub. Users provide a dataset and configure a few training parameters such as number of epochs and learning rate.
After training completes, the model is prepared for deployment by compiling it for the target hardware. This compilation step is handled automatically in both the command line and graphical tools.
Find the right model for your needs
Multiple models are available for each supported device. There is a balance between accuracy and speed, larger models will be more accurate but run more slowly. Find the right model for your needs.
Deploying your model
Model deployment requires the model be compiled beforehand to optimize it for the target hardware. With tools like Edge AI Studio, compilation is automatic. Otherwise, compiling models will require a separate step through software packages like edgeai-tidl-tools on the TI GitHub.
Model artifacts are deployed through runtimes like ONNX Runtime, tensorflow-lite, and TVM using TI Deep Learning (TIDL) as the hardware backend for acceleration. Models trained with TI tooling will train in PyTorch and export ONNX models for use with ONNX Runtime.
Deploying the model into an end-to-end vision application is accomplished with edgeai-gst-apps, which composes the pipeline with multiple stages of hardware acceleration for pre-processing and post-processing the image, in addition to accelerating the AI model itself.
Choosing the right device for you
TI Processors are benchmarked on a wide variety of models, and this data is available in Edge AI Studio. Multiple object detection models are available and can be run on devices with and without an NPU, although the performance difference shows that real-time processing strongly benefits from an NPU.
| Product number | Processing core | NPU available (TOPS) | Object detection benchmarks | |
|---|---|---|---|---|
|
YOLOX-Nano(416x416) latency [ms] |
SSD-Mobilenetv2 (512x512) Latency [ms] |
|||
| AM625 | 4x Arm® | No | 321ms | 705ms |
| AM62A7 | 4x Arm® | Y (2 TOPS) | 8.69ms | 14.51ms |
| TDA4VE-Q1 | 2x Arm® | Y (8 TOPS) | 4.23ms | 5.46ms |
All the hardware, software and resources you’ll need to get started
Hardware
SK-AM62A-LP
The AM62A is the lowest-cost AI-accelerated device in the AM6xA family, and is best suited for evaluation. A generic USB camera or webcam can be used for image capture and model evaluation on live data
Software & development tools
PROCESSOR-SDK-LINUX-AM62A
The Edge AI processor SDK is Linux-based and includes the necessary software components to run a compiled model with hardware acceleration. Other Edge AI accelerated processors may be substituted for AM62A
CCStudio™ Edge AI Studio
This tool contains tools for training, compiling and deploying a model to TI edge AI processors. A model selection tool is available to view pre-generated benchmarks of popular models.
Command-Line tools
Tools for Micro Processor devices with Linux and TIDL support. TI's edge AI solution simplifies the whole product life cycle of DNN development and deployment by providing a rich set of tools and optimized libraries.
TIDL tools
Find basic TI Deep Learning examples and compile custom model architectures.
Edgeai-gst-apps
Generate end-to-end applications on the Edge AI accelerated MPU with gstreamer using simple YAML configuration files
Supporting resources
The starting guide shows developers which aspects they need to consider for building an Edge AI application on TI Processors, and the tools they may require to do so