microsoft/onnxruntime

概览

解决什么问题

UAV developers need a single inference engine to deploy trained ONNX models across diverse onboard compute (CPU, GPU, NPU) without rewriting inference code for each backend.

主要用例

• Onboard object detection/tracking (e.g., YOLO, DETR exported to ONNX)
• Real-time segmentation for landing zone detection
• Obstacle avoidance inference pipelines
• Model serving in simulation-in-the-loop (AirSim, Gazebo) with hardware acceleration

次要用例

• Benchmarking model latency across execution providers
• Converting scikit-learn pipelines to ONNX for edge deployment
• Federated learning model inference on swarm nodes

何时使用

考虑使用

• Model is already in ONNX format or can be exported reliably
• Target hardware has a supported execution provider (CPU, CUDA, TensorRT, OpenVINO, QNN, CoreML, etc.)
• Deterministic latency and memory footprint are required
• License (MIT) permits commercial use without copyleft concerns

采用前需验证

• ? Execution provider availability and maturity for specific SoC (e.g., QNN for Snapdragon, TensorRT for Jetson)
• ? Operator coverage for model opset version
• ? Dynamic vs static quantization support for target provider
• ? Threading/affinity configuration for real-time OS integration
• ? Binary size impact on constrained flash/storage

从这里开始

采用检查清单

• 需验证 Execution provider availability and maturity for specific SoC (e.g., QNN for Snapdragon, TensorRT for Jetson)
• 需验证 Operator coverage for model opset version
• 需验证 Dynamic vs static quantization support for target provider
• 需验证 Threading/affinity configuration for real-time OS integration
• 需验证 Binary size impact on constrained flash/storage

以上检查项只有在官方来源确认后才能标记为“已验证”；无法确认的保持未验证。

已知限制与未知项

已知限制

• No built-in model training; inference-only
• Provider-specific operator support gaps (check compatibility per model)
• Large binary size when bundling multiple execution providers
• Real-time determinism depends on provider and OS scheduling
• NPU/accelerator support varies by vendor SDK version

替代与相关工具

相关工具

• tensorrt — 集成

如何使用？

从下方记录的入口开始，然后对照技术清单进行验证。

• 官方网站 https://onnxruntime.ai
• 仓库 https://github.com/microsoft/onnxruntime
• 文档 https://onnxruntime.ai

技术清单

• 通过 许可证已识别 已记录: MIT
• 通过 维护信号 活跃维护
• 通过 验证状态 官方确认
• 通过 已附加来源证据 1 个来源记录
• 需复核 已记录最新版本 未记录

官方链接

元数据与治理

数据集事实

以上事实仅来自官方数据集卡，未确认字段保持未知。

相关资源与依赖

• tensorrt — integrates with (confirmed)

近期动态

相关知识

下方相关指南/合集/对比的标题与摘要暂保留英文，待编辑翻译复核；事实字段与链接已验证且保持不变。

指南

• Deploying AI Inference on UAV Edge Compute — A repeatable workflow for choosing and validating an inference runtime for an onboard UAV computer.

合集

• 边缘AI感知入门套件 — 一套用于在边缘部署工作之前对空中感知模型进行原型验证的起步技术栈。