human-extractor_skill

Overview

This skill is a GPU-accelerated pipeline for detecting, tracking, and classifying humans in dashcam MP4 footage. It extracts per-person crops, optional CLIP-based head-covering scores, and writes a unified output directory with a comprehensive INDEX.csv for investigative workflows. The pipeline is optimized for NVDEC decoding, batched YOLOv8 detection, ByteTrack tracking, ROI extraction, and async WebP encoding.

How this skill works

The pipeline decodes videos (NVDEC if available), batches frames to YOLOv8 for person detection, then runs ByteTrack for multi-object tracking. Crops are GPU-extracted and optionally classified by CLIP for head-covering confidence. A deduplication filter (SSIM + rate caps) prunes redundant crops and an async I/O pool encodes WebP files and writes shard indices that are merged into INDEX.csv.

When to use it

• Extract and archive human appearances from dashcam or bodycam MP4 recordings.
• Perform investigative analysis where per-person crops and timestamps are required.
• Detect and filter people wearing head coverings using optional CLIP classification.
• Process large video batches with GPU acceleration to maximize throughput.
• Resume interrupted runs or produce auditable outputs with SHA1-backed INDEX entries.

Best practices

• Ensure CUDA and NVDEC drivers are installed to get the best throughput.
• Start with defaults (yolo_batch=64, clip_batch=384) then tune if OOM occurs.
• Enable save_full_frame only when annotated frame context is necessary.
• Use deduplication to limit storage growth (ssim=0.92, rate cap=12/min/track).
• Run pre-checks: model files present, output dir writable, torch.cuda available.

Example use cases

• Bulk-extract humans from a week of parking-lot dashcam footage for case review.
• Scan specified dates and save one annotated full-frame per timestamp for timeline assembly.
• Filter detections to only those with head coverings above an 80% CLIP threshold.
• Low-memory processing: disable CLIP, reduce yolo_batch, and turn off full-frame saves.
• Parallelize date-level jobs across workers to shorten wall-clock processing time.