overview
The i.MX93 Applications Processor delivers efficient, edge-ready machine learning and AI capabilities through a set of easy-to-use, pre-integrated demos included in the Linux Board Support Package (BSP). These demos make it simpler for developers—including those with limited embedded systems experience—to explore advanced ML and AI tasks. With smooth integration into Yocto-based builds and support for compatible evaluation kits, developers can quickly run and test key features like neural network acceleration and real-time inference, unlocking the full potential of the i.MX93 platform.
AI/ML Applications Supported by i.MX93 Processor
The i.MX93 processor supports a wide range of AI/ML applications including ML Gateway, Pose Estimation, Object Detection, Image Classification, LP Baby Cry Detection, Driver Monitoring System (DMS), Selfie Segmenter, i.MX Smart Fitness, Smart Kitchen, and more.
Image Classification Demo
This demo requires a mouse, display, camera (optional), and internet connection. It runs a GStreamer pipeline that captures video from either a camera or a video file and performs real-time image classification on each frame. The demo uses a pretrained, quantized MobileNetV1 model based on TensorFlow Lite, trained on a wide range of objects from the ImageNet dataset. Classification results are overlaid directly on the video, allowing users to see what objects the model recognizes in real time.
Object Detection
This demo performs real-time object detection using a quantized MobileNet SSD V2 model, trained on the COCO dataset. It processes video input from a camera or file, performing inference on each frame to detect and classify objects. The results, including bounding boxes and class labels, are displayed in real time. Implemented with a GStreamer pipeline, this setup is ideal for applications requiring efficient and real-time object detection.
Pose Detection with MoveNet Lightning
Real-time pose detection is made seamless with MoveNet Lightning. Using just a mouse, display, camera, and internet connection, you can set up a GStreamer pipeline to capture video from either a camera or file. Each frame is processed through a pretrained quantized MoveNet Single Pose Lightning TensorFlow Lite (TFLite) model, which accurately detects 17 key body points. These points are then overlaid onto the video feed in real-time, enhancing applications such as fitness tracking, interactive gaming, and gesture-based interfaces with efficient performance.
Driver Monitoring System (DMS) with GoPoint on i.MX 93
The Driver Monitoring System (DMS) is an essential safety feature in today’s vehicles, aimed at detecting driver fatigue, distraction, and inattention. Powered by the Neural Processing Unit (NPU) on the i.MX 93 platform, the system ensures efficient, real-time performance. The GoPoint application enables seamless integration, utilizing a camera interface to continuously monitor driver behavior with high accuracy.
Key Features of DMS on i.MX 93
Driver Status Indicators:
Provides real-time feedback on overall driver state with five specific behavior indicators—distraction, drowsiness, yawning, smoking, and phone usage.
Face Detection:
Identifies the driver’s face via live camera input, with a bounding box overlay for visual confirmation.
Alert Mechanism:
Triggers penalty scores and status updates when unsafe behaviors are detected or when the driver’s face is not visible.
Advanced Behavior Detection using ML Models:
- Distraction: Face not oriented forward
- Drowsiness: Eyes closed (non-blinking)
- Yawning: Mouth open
- Smoking: Cigarette presence detected
- Phone Usage: Mobile phone in use
By combining these features, the i.MX 93-based DMS offers a robust and efficient solution for enhancing driver safety through proactive, edge-based monitoring and intelligent alerts.
i.MX 93 Selfie Segmenter – Real-Time Background Separation
The i.MX 93 Selfie Segmenter leverages the onboard Neural Processing Unit (NPU) and NNStreamer to deliver fast, real-time person-background separation. Powered by a lightweight MobileNet V3 model with added decoder layers, the demo is available in two versions:
- 256×256 (General) – optimized for higher accuracy
- 144×256 (Landscape) – tuned for faster performance
Built using GStreamer and NNStreamer, the solution utilizes PXP hardware acceleration on i.MX 93 (and GPU on i.MX 8M Plus) for efficient video pre- and post-processing. This enables smooth, low-latency background substitution ideal for applications like:
- Video conferencing
- Augmented reality (AR)
- Photo enhancement
The result is a responsive and visually seamless user experience—all processed directly on the edge, without relying on the cloud.
Background Substitution Mode:
Enables dynamic replacement of the live background with a predefined or user-selected image, while overlaying real-time performance metrics—such as average FPS and inference latency (IPS)—in the bottom-left corner.
Key features of iW-RainboW-G50M
On Module Feature:
- i.MX 93 CPU with Dual Arm Cortex-A55 @1.7GHz
- NPU with up to 0.5 TOPS Neural Network Performance
- 2GB LPDDR4/4X Memory
- 16GB eMMC Flash expandable up to 128GB
- Wi-Fi 6 and Bluetooth 5.3 Connectivity
- -40°C to +85°C Industrial Grade
- 82mm x 50mm (SMARC v2.1.1 compatible SOM)
i.MX 93 SMARC System on Module:
Conclusion:
The i.MX93 Applications Processor is a versatile, power-efficient platform for deploying AI and ML at the edge, enabling real-time, on-device intelligence without reliance on the cloud. With robust support for applications such as image classification, object and pose detection, driver monitoring, and selfie segmentation, the platform is backed by an integrated Neural Processing Unit (NPU) and optimized TensorFlow Lite models.
iWave offers production-ready System on Modules (SoMs) and evaluation platforms based on the i.MX93, accelerating development across industries including automotive, healthcare, smart home, and fitness. Explore more at: www.iwave-global.com
