Baidu Yijian Vision Agent

What Problem It Addresses

Surveillance images and live scenes often need decisions such as person falls, vehicle crossings, occupied zones, or OCR. Raw frames alone are not enough; detection scope, counting rules, and reviewable output must be structured. This skill connects Baidu Yijian vision detection capabilities into a local workflow, turning image or video frames into inspectable structured results.

How It Works

• Skill selection and registration: reads the platform skill list and matches user intent, such as person detection, vehicle recognition, OCR, pose estimation, or tracking.
• Detection invocation: uses the YIJIAN_API_KEY environment variable to call a registered skill with an image or video frame; scripts write JSON to stdout and errors to stderr.
• Area constraints: defines an ROI electronic fence to limit detection scope, or a tripwire to count crossing events.
• Visualization: draws bounding boxes, generates grid references, and previews ROI/tripwire placement for coordinate confirmation.

Limits and Caveats

It requires access to the Baidu Yijian platform and a valid YIJIAN_API_KEY. It is intended for scenarios that already rely on vision detection services, not for deploying models, storing business data, or replacing a full monitoring backend; video workloads usually require frame extraction before per-frame processing.