C++ edge video AI you wire in a browser

What it does

CosmoEdge is a C++17 engine for production video analytics on edge devices. It ingests multi-channel video, runs detection and tracking through hardware-accelerated pipelines, renders real-time OSD overlays in the browser, and pushes structured events over MQTT or HTTP. A web console lets you compose pipelines visually—connecting cameras, models, post-processing nodes, alarm rules, and outputs—while the runtime handles decoding, inference scheduling, and memory pools on the device.

The interesting bit

The project treats vision-language models as asynchronous pipeline citizens rather than afterthoughts: you can drop in a GroundingDINO node for open-vocabulary detection from text prompts, or a Qwen3 VLM node that answers YES/NO state-judgment questions, running either on-device or against an OpenAI-compatible endpoint. That lets long-tail detection and visual reasoning share the same pipeline graph as conventional YOLO tracking and line-crossing analytics.

Key highlights

• C++17 runtime with hardware decoding and NPU integration, targeting Sophon BM1688 and x86 Linux/Windows.
• Browser-based visual pipeline editor for wiring video sources, AI models, OSD, alarms, and data exports.
• Async VLM and GroundingDINO nodes that operate alongside real-time CV pipelines without blocking the main flow.
• Real-time OSD with semantic coloring, zone overlays, and a debug view showing raw detections and track IDs.
• Model repository supporting ONNX and bmodel formats, with hot-swap workflows and version management.

Caveats

• x86 mode is explicitly a lower-throughput developer preview; production performance assumes Sophon BM1688 NPU hardware.
• On-device VLM support is tied to specific Qwen3 and Qwen3.5 variants, and the certified CosmoEdge-VL-Judge-0.8B package is a separate device-specific deliverable.
• The README documents workflows and features extensively, but leaves the internal C++ architecture and inter-process communication between runtime and console largely unexplained.

Verdict

Worth a look if you are building appliance-style edge video systems on Sophon hardware and want a unified runtime, pipeline UI, and data export layer in one deployable stack. Give it a pass if your deployment target is primarily cloud or if your edge fleet lacks BM1688 NPUs, since x86 mode is explicitly for development only.