Human Pose Recognition

Problem

In security review, elder care, or monitoring for people living alone, deciding whether a person has fallen or has remained in an abnormal posture often requires manual video review. This skill turns posture assessment into a queryable, structured result: given an image, video, or public URL, it classifies common postures such as standing, sitting, lying down, bending, raising hands, running, and falling, and highlights abnormal limb postures or fall events.

Core Capabilities and Flow

• Posture classification: uses skeletal keypoint detection and temporal action analysis to distinguish normal activity from abnormal behavior such as falls or rigidity.
• Fall alerts: --alert-fall is enabled by default, so detected falls or prolonged abnormal postures can trigger warning notes.
• Structured output: returns detection results, risk notes, suggestions, and report links, with verbosity options such as basic, standard, and json.
• Historical report lookup: --list reads the report list from the cloud interface and can filter by start date, instead of relying on local conversation records.

Typical usage provides either an --input local video path or a --url video address. If the input is an attachment, the skill first saves it as a local file. Identity association is handled internally, so users do not need to provide usernames, phone numbers, or other identity parameters; the response focuses on progress, results, and report links.

Boundaries and Notes

• Supported video formats include mp4, avi, and mov, with a maximum size of 10MB per file.
• Network URLs are treated as public addresses and downloaded by the API service, so local pre-download is not required.
• Results are intended as safety-monitoring references only; emergency situations still require immediate human response.
• Use the skill's built-in scripts and provided references; avoid generating replacement scripts, and read API documentation only when interface specs or error codes are needed.