A growing number of industrial applications share the same fundamental requirement: a system that can identify what it’s looking at and know exactly how far away it is. A warehouse robot needs to recognize a pallet and measure the distance to it. An inspection camera on a production line needs to classify a defect and locate it in three-dimensional space. An autonomous vehicle needs to detect a pedestrian and determine their precise position relative to the chassis.

Until recently, building a system that could both identify objects and measure their 3D position at the edge meant combining together multiple boards, dedicated FPGAs, and external depth sensors: high cost, large form factor, and specialized development effort. What changes when AI inference and 3D reconstruction converge on a single, low-power embedded System-on-Chip?

The result is a complete AI and 3D vision pipeline running on the Renesas RZ/V2H, a vision-focused embedded SoC from Renesas designed for edge applications, combining real-time object detection with metric depth measurement, all executing on a single chip. Deep Vision built this pipeline as a ready-to-run starting point for teams exploring smart camera and embedded 3D vision applications. It’s available today for hands-on evaluation, with details on how to get it at the end of this post.

Most embedded SoCs targeting vision and AI applications offer a neural processing unit (NPU) as their primary accelerator. The RZ/V2H does too, its DRP-AI accelerator handles neural network inference for tasks like object detection and classification. But NPU-only chips hit a wall when the application demands more than AI.

Stereo depth computation is a case in point. The Semi-Global Matching (SGM) algorithm at the core of classical stereo vision is computationally demanding but poorly suited to NPU architectures, it relies on dynamic programming and cost aggregation patterns that don’t map efficiently to matrix-multiply hardware. Running SGM on the CPU is possible but slow and resource-hungry, leaving little headroom for anything else.

The RZ/V2H addresses this with a second accelerator: the Dynamically Reconfigurable Processor (DRP). The DRP is a programmable hardware fabric optimized for the kind of structured, non-neural compute that stereo matching requires. This is what makes the AI and 3D vision pipeline possible on a single chip, the DRP-AI handles object detection while the DRP simultaneously handles depth computation, and neither competes for the same resource.

Equally important is the RZ/V2H’s native camera connectivity. The SoC provides four MIPI CSI-2 inputs, enough to connect a stereo camera pair (and more) without requiring external multiplexers or bridge chips. Stereo vision demands synchronized image pairs from two cameras; having dedicated, on-chip camera interfaces simplifies the hardware design and reduces both cost and board area.

Without this specific combination, NPU for AI inference, DRP for classical vision acceleration, and multiple MIPI inputs for native multi-camera support, the full AI and 3D vision pipeline cannot run on a single embedded SoC. This is what distinguishes the RZ/V2H from the broader field of edge AI processors. For a detailed technical evaluation of the RZ/V2H’s architecture, accelerators, and individual subsystem benchmarks, see our previously published post: Renesas RZ/V2H Embedded Vision Platform: A Comprehensive Evaluation.

The pipeline takes the bounding boxes produced by the DRP-AI’s YOLO model and maps them into the DRP’s disparity map. For each detected object, the system reads the corresponding disparity values within the bounding box and converts them to a metric distance using the stereo camera’s known baseline and focal length. The result: every detection is annotated with its real-world distance from the camera.

The depth data is also rendered as an interactive 3D point cloud on the onboard GPU, allowing the scene to be explored directly on a connected display: this makes the spatial output tangible and inspectable without any external compute or visualization software.

Combining detection with calibrated distance is what turns a camera from a classification sensor into a spatial measurement tool. Knowing what an object is and where it is, in real-world units, is the foundation for applications that need to act on the physical world: collision avoidance, pick-and-place guidance, spatial occupancy mapping, or proximity-triggered workflows.