AI Wildlife Camera

Automated wildlife detection pipeline for infrared trail camera footage. Version 2.2 — pre-execution interaction + configurable frame density.

What's New in v2.2

Workflow Overview

5-phase pipeline for batch-processing trail camera videos:

关键变更（v2.2）：所有交互确认在 Phase 0 完成，确认后才执行任何文件操作。

Prerequisites

• FFmpeg with ffprobe (ffmpeg.exe / ffprobe.exe on Windows)
• Python 3.8+ with openpyxl
• NVIDIA GPU (optional, speeds up if vision model runs locally — currently uses agent vision)

Configuration

Edit scripts/inventory.py and scripts/extract_frames.py top CONFIG section:

FFMPEG_BIN = r"C:pathtoffmpegbin"      # Windows
# FFMPEG_BIN = "/usr/bin"                    # Linux/macOS

INPUT_DIR = r"C:TrailCameraVideos"        # Your footage folder
OUTPUT_DIR = r"C:TrailCameraOutput"        # Results folder

Phase 0: 预执行交互确认（强制）

⚠️ 执行任何文件操作前，Agent 必须先完成以下交互，获得用户明确确认后方可继续。

Step 0a: 询问相机安装地点

Prompt the user:

"🦌 即将开始野生动物视觉识别。请指定红外相机安装地点（至少精确到省/地区，如'中国云南省'），这会用于修正物种识别结果。默认'中国'。"

Store location in output/location.txt (single line, e.g. 中国云南省高黎贡山).

Step 0b: 说明当前模型并询问用户选择

Prompt the user with full transparency:

"🧠 当前可用的视觉识别模型：

• A) Qwen-VL-Plus API（默认，通过阿里云百炼接口，批量处理，支持物种对比推理、跨帧去重、断点续传）
• B) Kimi 内置视觉（我逐帧查看，适合小批量或 API 不可用时）
• C) 其他 API（GPT-4o / Claude / Gemini，需要你自行提供 API key）

请选 A/B/C，或告诉我你的偏好。如选 C，请提供 API key 和模型名称。"

根据用户选择配置对应脚本：

• 选 A → 确认 analyze_api.py 中已有阿里云 key，可直接使用
• 选 B → 进入 Agent 手动 review 模式
• 选 C → 用户提供 key 后写入 analyze_api.py CONFIG，或创建新的分析脚本

Step 0c: 询问帧密度

Prompt the user:

"📐 帧密度选择（影响识别精度和 token / API 消耗）：

• 高：当前设置（三段式共 8-28 帧/视频，精度最高，token 消耗最大）
• 中：调减 50%（三段式共 4-14 帧/视频，平衡精度与成本）
• 低：调减 75%（三段式共 2-7 帧/视频，成本最低，适合快速初筛）

请选 高/中/低。"

Store density selection in output/frame_density.txt (single line: high / medium / low).

Frame density scaling rules:

最低保障：每个 zone 至少提取 1 帧，确保覆盖视频前中后三段。

Step 0d: 最终确认

Prompt the user:

"📋 确认信息：

• 地点：[用户提供的地点]
• 模型：[用户选择的模型]
• 帧密度：[高/中/低]
• 待处理视频：[INPUT_DIR 路径]

确认无误后回复'开始'，我将执行扫描→帧提取（按选的密度）→视觉识别→导出报告。"

收到用户明确回复（如"开始"/"确认"/"跑吧"）后，方可进入 Phase 1。

Phase 1: Inventory (Auto)

python scripts/inventory.py

Scans INPUT_DIR recursively for video files (.mp4, .mov, .avi, .mkv, .m4v, .mpg, .mpeg).

Extracts per video:

• 原始文件名 — filename
• 拍摄时间 — Parsed from filename patterns, then ffprobe creation_time, fallback to file mtime
• 视频时长 — Duration in seconds
• 分辨率 — Width × Height
• 文件大小 — MB
• 编码格式 — Video codec

Outputs:

• output/inventory.json — Raw data
• output/inventory.xlsx — Excel preview (Phase 1 data only)

Date Parsing Priority

1. Filename patterns: IMG_YYYYMMDD_HHMMSS, YYYY-MM-DD_HH-MM-SS, YYYYMMDD_HHMMSS, YYYY_MM_DD_HH_MM_SS
2. ffprobe creation_time / date tag
3. File modification time (fallback)

Phase 2: Frame Extraction (Auto) — Three-Zone + User Density

python scripts/extract_frames.py

Reads output/inventory.json and output/frame_density.txt, extracts frames per video using three-zone strategy with user-selected density.

Frame naming: Flat structure — output/frames/PIRT0001_frame_001.jpg, PIRT0001_frame_002.jpg, etc.

Frame extraction reads frame_density.txt to determine scaling factor:

• high → 100% of base frame count (8-28 frames)
• medium → 50% of base frame count (4-14 frames)
• low → 25% of base frame count (2-7 frames)

Base frame count (high density):

Rationale for three-zone split: Trail camera videos are triggered by motion, but animals may enter at start, linger mid-video, or exit at the end. Three-zone coverage maximizes detection probability across the full clip.

Frame resolution: 800px width (up from 640px) for better small animal detail.

Frame quality: JPEG quality=2 (high).

Phase 3: Vision Analysis + Correction + Write Results

Step 3a: Run Vision Analysis

根据用户在 Phase 0b 的选择，执行对应的视觉识别：

Option A — API Batch Mode (Qwen-VL)

python scripts/analyze_api.py

Sends frames per video to Qwen-VL-Plus API (frame count depends on density selected in Phase 0c).

Option B — Agent Manual Review

Agent views extracted frames using read tool on image files and records per-video summary.

Option C — Other API

Use user-provided API key and model.

Step 3b: Apply Location-Based Correction

After raw vision results are in, read output/location.txt and apply correction rules:

• 不在该地区分布的物种 → 排除或降置信度
• 该地区常见物种 → 提升置信度
• 候鸟/迁徙种 → 标注季节性
• 中国家养动物 → 与野生动物区分
• 入侵物种 → 特别标注
• 易混淆物种对 → 用区分特征修正。参考 references/wildlife_guide.md 中的地区物种参考和形态特征描述，对相似物种进行排除法推理。修正时不硬编码具体物种对，而是根据实际检出结果动态比对地区常见物种的形态特征（体型、毛色、尾型、行为模式等）。

Step 3c: Write Results to vision_analysis.json

After applying location-based correction, write the final results to output/vision_analysis.json:

{
  "location": "中国云南省高黎贡山",
  "model_used": "qwen-vl-plus",
  "frame_density": "medium",
  "correction_applied": true,
  "videos": [
    {
      "filename": "RCNX0001.AVI",
      "has_human": false,
      "has_wildlife": true,
      "species_detected": ["野猪"],
      "individual_count": {"野猪": 2},
      "confidence": "high",
      "notes": "夜间拍摄，成年个体带幼崽，从左侧进入画面"
    }
  ]
}

Field reference:

Writing command (for Agent or script):

import json

vision_analysis = {
    "location": location,  # from location.txt
    "model_used": model_name,  # e.g. "qwen-vl-plus" or "kimi-vision"
    "frame_density": density,  # from frame_density.txt
    "correction_applied": True,
    "videos": corrected_results  # list of dicts
}

with open("output/vision_analysis.json", "w", encoding="utf-8") as f:
    json.dump(vision_analysis, f, ensure_ascii=False, indent=2)

Phase 4: Export to Excel (Auto)

python scripts/export_excel.py

Reads inventory.json + vision_analysis.json, merges data, writes structured Excel:

Output: output/wildlife_report.xlsx

Color coding:

• 🟢 浅绿 — 有野生动物
• 🟠 浅橙 — 有人类
• 🔴 浅红 — 读取错误行

Batch Processing Tips

• For >50 videos: split into folders of 20–30 videos to manage frame review load
• Frame extraction can run overnight; vision review can resume per-folder
• Use --fps 1 if you want 1 frame per second (modify extract_frames.py CONFIG)

Known Limitations

• Audio is ignored — no acoustic species identification
• API dependency — Qwen-VL API requires valid key and network; fallback to agent manual review when unavailable
• Night/IR footage — low-light frames may still reduce accuracy; infrared-trained models TBD for v3
• Small animals — v2 improved with 800px frames and three-zone sampling, but very distant rodents/birds may still be missed
• Cross-model validation — v2 uses single API model; v3 planned: dual-model consensus (Kimi + Qwen-VL) with disagreement flagging