Hi, thanks for starting this discussion. These are exactly the trade-offs we run into on real deployments, so here are some thoughts on the first three questions:
1. Which ARM platforms suit real-time transportation applications?
It mainly depends on the AI workload and power budget:
- Telematics / gateway (GPS, CAN, 4G/5G, light analytics): NXP i.MX 8M Plus / i.MX 93 class SoCs offer low power consumption, and the i.MX 8M Plus includes an NPU for small vision models.
- Multi-camera / driver monitoring: You’ll need more AI compute, e.g. NXP i.MX 95, Qualcomm QCS6490 / Dragonwing IQ9 series, NVIDIA Jetson Orin, or an Arm platform paired with an AI accelerator module.
- Beyond the SoC, check for wide-temperature operation, wide-range DC input, CAN bus, GNSS, and any in-vehicle certifications required in your market. In practice these often matter as much as raw compute.
Hands-on examples: Edge AI Video Assistant on AOM-2521 (i.MX 95 + Ara240) and Advantech Arm platform + Hailo AI accelerator.
2. Is Embedded Linux a better choice for edge devices?
For most edge AI gateways, yes. You get broad driver support, container support (Docker/Podman), ROS 2, and mature AI runtimes, and with Yocto you can build a minimal, reproducible image. For hard real-time requirements, Linux is often combined with an RTOS (e.g. on a Cortex-M core or a separate MCU), or a PREEMPT_RT kernel is used.
See: Get started with ROS on Yocto
3. How to deploy lightweight AI models on constrained hardware?
- Quantize to INT8
- Use the vendor NPU toolchain (eIQ for NXP, QNN/SNPE for Qualcomm, TensorRT for Jetson) instead of running on the CPU
- Choose edge-friendly models (e.g. YOLO n/s variants, MobileNet)
Step-by-step guides: Deploy AI/ML models on QCS6490 and Object detection on AOM-2721 with Edge Impulse.
More how-to articles are in our Knowledge Base. If you’re working on a specific platform or use case, feel free to share the details. It would be great to compare notes here.