How dedicated AI hardware transforms automotive quality control and enables multi-camera, low-latency edge analytics
In recent years, AI-enabled vision systems have fundamentally transformed the landscape of automotive vehicle inspection. When combined with advanced AI models that can evaluate thousands of images per second, modern camera systems can now detect production inaccuracies down to minute deviations in a vehicle’s paint job. With inspection happening in real time during production, these advances in machine vision have driven significant improvement in production quality across automotive manufacturing.
Similarly, these systems can be used by automotive insurance companies to dramatically reduce processing times for claims. Here, AI image recognition can quickly detect, assess and categorize vehicle damage from images uploaded by the customer, removing the need for manual checking. Repair workshops can also benefit from intelligent image analysis to precisely identify damages and calculate repair costs. This promotes shorter vehicle downtime and more transparent cost estimates, resulting in higher customer satisfaction.
Processing AI Models in Real Time at the Edge
Edge computing platforms play a critical role in AI vision for modern production environments. Unlike AI processing in the cloud, these platforms operate where the data is generated—directly at the camera or image sensor—and leverage AI accelerators for real-time inferencing against streaming video of hundreds or thousands of images per second from high-speed cameras.
Such video processing pipelines take advantage of a combination of general-purpose central processing unit (CPU), graphic processing unit (GPU), image sensor processor (ISP), vision processing unit (VPU), neural processing unit (NPU) or AI accelerator hardware. The various units are selected based on the operations required of the application and the most efficient way to accomplish these operations.
Typically, raw video or image data first makes its way to local memory, where it is loaded by a CPU, VPU, or ISP that pre-processes it to reduce noise or batch frames. This pre-processed data is then written back to local memory before being loaded by an onboard accelerator that is optimized for running defect detection, object recognition, or other analysis using an AI model or algorithm. Results of that inference are written back into shared memory, where they are post-processed by a CPU, VPU, or ISP before being output to a display, control loop, or other subsystem of the larger automotive quality inspection system.
Figure 1
Figure 1. The Axelera Metis AIPU onboard the SOM-COMe-BT6-RK3588 provides up to 214 TOPS of computational AI performance—enough to process eight HD camera feeds simultaneously and still perform optical character recognition and object detection.
The combination of an AI accelerator and tightly designed onboard memory architecture allow for low-latency processing of multiple video streams simultaneously on a small form factor edge device. One such platform is the SOM-COMe-BT6-RK3588, a COM Express Type 6 Basic module that is well suited for compact AI vision systems thanks to its 120 × 95 mm form factor. This solution is based on the powerful Rockchip RK3588 CPU, which features four Arm Cortex-A76 and four Cortex-A55 application cores alongside three Cortex-M0 cores for control tasks. Supported by up to 32 GB LPDDR5-3200 memory, the RK3588 also integrates an Arm Mali GPU, an NPU, and a VPU, making it a formidable system for advanced video processing workloads.
However, the highlight of the SOM-COMe-BT6-RK3588 is the Axelera AI Metis AIPU (AI processing unit). In benchmark tests, this specialized hardware accelerator has been shown to deliver up to 214 TOPS of AI performance, making it among the best in its class. Achieving approximately 15 TOPS per watt in isolation, the Metis AIPU is also highly energy efficient, making it possible to run even multi-stream AI workloads—up to 24 parallel camera streams with real-time object detection, for example—on one resource-optimized edge system.
As part of SECO’s COM Express module, the Metis AIPU grants developers up to 120 TOPS of AI performance in real-world settings. The module’s onboard Rockchip RK3588 CPU supports up to four cameras—and potentially even more via channel virtualization and aggregation using technologies like MIPI CSI-2 virtual channels (VCs). Now, developers can build advanced vision systems for automotive inspection applications by utilizing the high-end AI processing capabilities of the Metis AIPU.
Harnessing AI Acceleration for Video Processing
To simplify the development of high-performance AI vision at the edge, Axelera AI’s Metis AIPUs are supported by its Voyager SDK. This user-friendly software stack automatically compiles, optimizes, and deploys the entire AI pipeline while also offering compatibility with popular AI development frameworks such as LiteRT (formerly TensorFlow Lite) and PyTorch.
As a result, developers can leverage the full potential of the Metis AIPU for a wide set of use cases. With Metis, for example, OEMs and systems integrators supplying automotive quality inspection systems can choose from a variety of different implementation options on the same foundational building blocks.
One such implementation involves using two external quad deserializers and MIPI-CSI VC to aggregate up to eight camera streams across each of the four lanes of the SOM-COMe-BT6-RK3588's dual onboard CSI connectors.
Axelera AI maintains an extensive GitHub library with many well-documented example projects like the 8x 1080p60, 4x 4K30, or 1x 8K30 demos highlighted below. These help developers get started in real-world settings.
Figure 2
Figure 2. The SOM-COMe-BT6-RK3588 supports a range of I/O that can be extended for specific use cases. In an automotive manufacturing quality inspection system, for example, two quad deserializers could be added to the COM Express Type 6 module’s companion carrier board and used with MIPI Virtual Channels (MIPI VC) to aggregate streams from up to eight cameras.
| Scenario | Estimated throughput & latency | Supporting Axelera demo |
|---|
| 8×1080p@60 FPS streams (object detection + light OCR) | Realtime throughput on the integrated ~120 TOPS Metis AIPU
Latency 20-40 ms per stream (decode/resize 6–12 ms; detector 6-15 ms; OCR 2–6 ms; post-processing 2-5 ms) | Metis is designed for many concurrent video streams; 1080p YOLOv8S fits comfortably within the 120 TOPS budget. |
| 4×4K@30 FPS streams with cascade (detector + segmentation) | Realtime with region of interest (ROI) cascades
Latency 35-70 ms, depending on ROI counts. | Use cascade pattern from fruit demo, cropping high-resolution ROIs for segmentation. |
| Single 8K@30 FPS surface inspection | Near real-time using tiling | 8K tiling demo processes hundreds of overlapping tiles per second; scaling AIPUs increases tile throughput. |
| Many short MIPI/USB cameras (e.g., 12×720p) | Throughput limited by I/O and host preprocessing; AIPU still has headroom | Metis can run over 16 concurrent video streams; the RK3588’s multiple CSI/USB controllers handle input. |
As shown in Table 1, the Metis AI accelerator can process anywhere from one to eight high-definition (HD) video feeds while running AI inference at or near real-time. It can also support 12 or more standard definition cameras while still retaining headroom for additional workload.
A Promising Roadmap for Edge-AI Inspection Systems
By integrating a dedicated AI hardware accelerator into its SOM-COMe-BT6-RK3588 module, SECO has shown developers that the COM Express standard is ready to deliver advanced machine vision performance to support next generation edge-based AI. This shift will enable new system form factors with greater throughput and higher processing efficiency, which will be essential as AI technologies further streamline automotive vehicle inspection tasks.
For more information on SECO edge computing platforms featuring or compatible with Axelera AI technology, visit seco.com or the SECO App Hub.