My home surveillance had stacked up hundreds of GB, mostly frames of nothing happening. Goal: keep the moments with people, cut everything else — fast, and fully automatic. Using YOLOv8n (OpenVINO, on a hybrid iGPU+NPU pool) for per-second person detection, I auto-compressed 1,677 directories / 79,027 files on D: — 502.5GB → 235.7GB (−53%) in 6.5 hours overnight, zero errors. This is the full battle log.

Env: Core Ultra 5 235H / Arc 140T (copy-cut) / Intel AI Boost NPU (detection) dual pool, GPU×4+NPU×3 / OpenVINO 2026.5.0.dev / yolov8n_openvino_model FP16 / ffmpeg 9.0.1 / Windows Task Scheduler. Tested 2026-09-12~13.

Why we defected from freezedetect to YOLO

We started with ffmpeg freezedetect / MOG2 / frame-diff 'motion detection', which collapsed against surveillance night footage: sensor noise made static frames look 'active', while real walk-bys came out as fragmented frame diffs. It can't tell 'static' from 'has people' — those are different problems entirely. The final ruling landed on keep-if-there-is-a-person, and only an object detector can make that call.

The three-stage pipeline

1. Detect (resident GPU×4+NPU×3 service): one frame per 3s through YOLOv8n → global activity-segment JSON (people_secs / intervals)
2. Cut (stream copy): map segments back to in-file seconds, ffmpeg -ss/-t per slice, -c:v copy = zero quality loss
3. Merge (concat): demuxer joins slices into the final 监控成片\目录名.mp4

From 15× to 127×: the speed-up ledger

This pipeline came from three rounds of iteration, each halving the time:

Two decisions mattered most: stream-copy cutting instead of re-encoding (44 slices in 2.8s vs 87s re-encode — 24× faster), and detecting a full hour of video in under 17s (hybrid pool + 3s sampling).

Why cut with copy, not re-encode

The source is HEVC — already highly compressed, so re-encoding is slow and lossy. With detection accurate to the second, just -c:v copy the matching packets and stitch — zero quality loss, negligible size cost. That's the core of 127× realtime.

AV sync and quality spot-checks

Final outputs spot-checked (HEVC 1080p@20 + AAC 16k mono): resolution/framerate/audio/timing all correct; AV offset 64ms (video start 0.064s / audio 0s) acceptable; frame at 820s has brightness 117 — real content, not black/green.

Full-disk deploy: task scheduler + resume

The scale is brutal: 1,677 dirs / 79,027 files / 502.5GB, D: is read-only and nothing is deleted. A raw background command gets killed; the fix is a Windows scheduled task:

:: 计划任务「监控压缩全量」→ launch_all.py（自包含：确保服务运行→调 batch_run）
:: batch_run.py D:\ --jobs 8 --interval 3 --pad 3 --skip-existing --log batch_log.txt
:: --skip-existing：跳过已完成目录，中断后重跑即断点续传（速率 ~6.1 目录/分钟）

Deployment traps: a scheduled task calling python directly failed (Last Result=1) from Chinese paths / D:\ serialization pollution; a cmd wrapper then broke on backslash-quote parsing. The fix was launch_all.py, a pure-Python launcher with args hardcoded (forward-slash D:/), and the task invoking a no-arg script — ran to completion.

Results

• 1,677/1,677 directories completed, zero errors; 1,554 finals (123 people-less dirs yielded none)
• Size 502.5GB → 235.7GB, −53%
• ~6.5 hours total (23:51 → 06:26), and the box stayed usable throughout
• All outputs land in Desktop 监控成片\目录名.mp4; temp slices auto-clean