Kami Fall Detection (Cloud)

Detect fall events from RTSP camera streams using the KamiClaw cloud API. No local GPU needed — all inference runs server-side.

How it works

1. Stage 1 — Change detection (local, CPU-only): Lightweight frame-differencing on 128×128 grayscale. Keeps a rolling pre-buffer (default 4s). When a significant temporal change is detected, triggers Stage 2.
2. Stage 2 — KamiClaw cloud analysis: Collects post-transition frames (default 4s), resizes the clip to 360×360, writes to a temp file, and sends it to the KamiClaw API (detectType: FALL, detectSubType: SK_FALL_DETECTION). If a fall is detected, outputs a JSON alarm line and optionally saves the clip. Then resets and returns to Stage 1.
3. Auto-reconnect: If the RTSP stream drops, the skill automatically retries connection every 5 seconds and resumes monitoring seamlessly.

API Key Setup Guide

This skill requires a KamiClaw API key to authenticate with the cloud service.

Step 1 — Register

Go to https://kamiclaw-skill.kamihome.com and create an account. Choose a plan that fits your usage.

Step 2 — Get your API key

After registration, copy your API key from the dashboard. It looks like:

sk_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Step 3 — Provide your key and RTSP URL

Tell the agent your API key and RTSP URL. The agent will automatically save them to config.json.

Agent instructions:

1. Before running this skill, read config.json and check if api_key and rtsp_url are both non-empty.
2. If either is empty, ask the user to provide the missing value(s). Do NOT proceed until both are provided.
   • If api_key is empty: "Please provide your KamiClaw API key. Register at https://kamiclaw-skill.kamihome.com to get one."
   • If rtsp_url is empty: "Please provide your RTSP camera URL (e.g. rtsp://192.168.1.100/live/stream) or a local video file path."
3. Also ask: "Would you like to save video clips when a fall is detected? (yes/no, default: yes)" — set save_alarm_clips to true or false accordingly.
4. Ask about Feishu notifications: "Would you like to receive fall alarms via Feishu? (yes/no)"
   • If yes: Guide user to create a Feishu group bot and get the webhook URL
   • Fill feishu_webhook_url in config.json with the webhook URL (e.g., https://open.feishu.cn/open-apis/bot/v2/hook/xxx)
   • If no: Leave feishu_webhook_url as empty string ""
5. Once the user provides all values, write them into config.json (read existing, merge, write back).
6. Then run the skill.

Do NOT ask the user to manually edit config.json. Always write it for them.

Installation

bash setup.sh

This will:

1. Detect system Python (python3 → python → auto-install)
2. Create .venv/ virtual environment
3. Install: requests, opencv-python-headless, numpy

No GPU or heavy ML libraries needed.

Running

.venv/bin/python fall_detect_cloud_skill.py \
    --rtsp_url rtsp://192.168.1.100/live/stream \
    --api_key sk_live_xxxxxxxx

Or set values in config.json and run without arguments:

.venv/bin/python fall_detect_cloud_skill.py

Configuration

All parameters are in config.json. The agent auto-fills this when the user provides values in chat.

Feishu Notification Setup (Webhook)

To enable automatic alarm notifications via Feishu Webhook:

Step 1: Create a Feishu Bot in your group

1. Open your Feishu group chat
2. Click group settings (⋮) → Add Bot
3. Choose Custom Bot
4. Set a name (e.g., "Fall Detection Alert")
5. Copy the Webhook URL (format: https://open.feishu.cn/open-apis/bot/v2/hook/xxx)

Step 2: Add to config.json

{
  "feishu_webhook_url": "https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx"
}

Step 3: Test the notification

.venv/bin/python fall_detect_cloud_skill.py

The system will send a test message to confirm the webhook works.

Agent instructions: When configuring this skill, ask the user if they want Feishu alarm notifications. If yes, guide them to create a Feishu group bot and provide the webhook URL.

CLI args and env vars still work as overrides (priority: CLI > env > config.json > defaults).

Agent: If api_key or rtsp_url is empty in config.json, you MUST ask the user to provide both before running. Never run the skill with either field empty.

Output Format (stdout JSON)

When alarm triggers:

{
  "alarm": true,
  "type": "fall",
  "fall_type": "active_falling",
  "num_persons": 1,
  "confidence": 0.92,
  "reason": "Person collapsed from standing to floor near bed",
  "frame": 87,
  "source": "rtsp://192.168.1.100/live/stream",
  "clip": "logs/clips/alarm_fall_20260429_143201.mp4"
}

Agent: When a fall alarm is detected (alarm: true), you MUST notify the user through the connected channel (e.g. chat, Feishu, WeChat) with the alarm details, including fall_type, confidence, reason, and clip path (if available).

When no alarm (run_time reached):

{
  "alarm": false,
  "type": null,
  "detail": "Run time limit reached, no fall detected",
  "frames_processed": 3600,
  "source": "rtsp://192.168.1.100/live/stream"
}

Exit Codes

The skill runs continuously and prints a JSON alarm line to stdout each time a fall is detected. It does NOT stop after an alarm — it resets and keeps monitoring. Stream disconnections are handled automatically via auto-reconnect.

Logs

All logs are saved in the logs/ directory:

alarms.jsonl format

Alarm triggered:

{"event": "alarm", "alarm": true, "type": "fall", "fall_type": "active_falling", "num_persons": 1, "confidence": 0.92, "reason": "...", "frame": 87, "source": "rtsp://..."}

Clip analyzed, no alarm:

{"event": "clip_analyzed", "frame": 120, "source": "rtsp://...", "fall_detected": false, "confidence": 0.97, "reason": "...", "_ts": "2026-04-27T14:33:15+0800"}

transitions.jsonl format

{"event": "transition", "frame": 83, "source": "rtsp://...", "_ts": "2026-04-27T14:31:58+0800"}

Analysis tips

# Count total alarms today
grep '"alarm"' logs/alarms.jsonl | grep "$(date +%Y-%m-%d)" | wc -l

# Count transitions vs alarms (false positive rate)
wc -l logs/transitions.jsonl logs/alarms.jsonl

# Extract all alarm events as CSV-friendly
cat logs/alarms.jsonl | python3 -c "
import sys, json
for line in sys.stdin:
    obj = json.loads(line)
    if obj.get('event') == 'alarm':
        print(f\"{obj['_ts']},{obj.get('type','')},{obj.get('fall_type','')},{obj.get('confidence',0)},{obj.get('frame',0)}\")
"

File Structure

kami-fall-detection-cloud/
├── SKILL.md                      # This file
├── config.json                   # All config in one place (api_key, rtsp_url, etc.)
├── fall_detect_cloud_skill.py    # Production skill entry point
├── fall_detect_cloud.py          # Development/debug version (with test mode)
├── setup.sh                      # Venv installer
├── requirements.txt              # Dependencies: requests, opencv-python-headless, numpy