OrangePi3588Media/docs/config_guide.md

16 KiB
Raw Blame History

配置文件编写指南

本文档说明当前项目主线配置的写法,重点覆盖两阶段鞋检测、劳保鞋颜色判断、告警节流,以及 RK3588 上与性能稳定性直接相关的参数。


1. 当前推荐架构

对车间劳保鞋场景,当前推荐链路是:

input_rtsp
  -> preprocess(rgb)
  -> ai_yolo(person only)
  -> tracker(person only)
  -> ai_shoe_det(dynamic roi)
  -> logic_gate(person_shoe_check)
  -> logic_gate(ppe_boots_check)
  -> osd
  -> preprocess(nv12)
  -> publish
  -> alarm

业务目标:

  • 先检测到人
  • 再只在脚部 ROI 检鞋
  • 只关心“检测到鞋后,颜色是否接近黑色劳保鞋”
  • 不再把“没穿鞋”作为主告警目标

2. 配置结构

当前项目常用的是 graph 模式:

{
  "queue": {
    "size": 8,
    "strategy": "drop_oldest"
  },
  "graphs": [
    {
      "name": "person_shoe_two_stage_workshoe_alarm",
      "executor": {
        "batch_size": 2,
        "run_budget": 8
      },
      "nodes": [ ... ],
      "edges": [ ... ]
    }
  ]
}

2.1 queue

{
  "queue": {
    "size": 8,
    "strategy": "drop_oldest"
  }
}

说明:

  • size:默认队列长度
  • strategy
    • drop_oldest:推荐,实时性最好
    • drop_newest
    • block

2.2 executor

{
  "executor": {
    "batch_size": 2,
    "run_budget": 8
  }
}

说明:

  • batch_size:执行器单次批量处理帧数
  • run_budget:单次调度预算

在 RK3588 上,如果出现“隔几秒卡一下”,通常先减小 run_budget,再看是否需要调 batch_size


3. 关键节点

3.1 input_rtsp

{
  "id": "in",
  "type": "input_rtsp",
  "role": "source",
  "enable": true,
  "url": "rtsp://10.0.0.49:8554/cam",
  "fps": 30,
  "width": 1920,
  "height": 1080,
  "use_ffmpeg": true,
  "use_mpp": false,
  "force_tcp": true
}

当前推荐:

  • 对会在固定画面卡顿的 RTSP 源,优先使用 use_ffmpeg: trueuse_mpp: false
  • 原因是部分源流在 ffmpeg demux + mpp decode 路径上会出现固定位置卡顿,而 VLC 直拉源流正常

3.2 preprocess

{
  "id": "pre_rgb",
  "type": "preprocess",
  "role": "filter",
  "enable": true,
  "dst_w": 1920,
  "dst_h": 1080,
  "dst_format": "rgb",
  "dst_packed": true,
  "resize_mode": "stretch",
  "rga_gate": "person_shoe_two_stage_workshoe_alarm",
  "use_rga": true
}

说明:

  • 前级 pre_rgb 准备共享的高分辨率 RGB 主帧
  • 后级 post 再转回 nv12 给编码器
  • use_rga: true 是 RK3588 上的推荐项

3.3 ai_yolo

{
  "id": "person_det",
  "type": "ai_yolo",
  "role": "filter",
  "enable": true,
  "infer_fps": 2,
  "infer_phase_ms": 0,
  "use_rga": true,
  "use_dma_input": true,
  "model_path": "./models/yolov8n-640.rknn",
  "model_version": "v8",
  "model_w": 640,
  "model_h": 640,
  "num_classes": 80,
  "conf": 0.35,
  "nms": 0.45,
  "class_filter": [0],
  "bbox_expand": {
    "enable": true,
    "class_id": 0,
    "left": 0.06,
    "right": 0.06,
    "top": 0.04,
    "bottom": 0.16
  }
}

当前用途:

  • 只做人检测前级
  • 不再直接承担鞋检测

关键参数:

参数 作用 当前建议
infer_fps 人体检测频率 2
infer_phase_ms 与鞋检错峰 0
class_filter 只保留 person [0]
bbox_expand.bottom 补偿漏脚 0.16 左右
use_rga / use_dma_input 输入性能优化 建议开启

3.4 tracker

{
  "id": "person_trk",
  "type": "tracker",
  "role": "filter",
  "enable": true,
  "mode": "bytetrack_lite",
  "per_class": true,
  "track_classes": [0],
  "high_th": 0.55,
  "low_th": 0.10,
  "iou_th": 0.3,
  "max_age_ms": 900,
  "min_hits": 2
}

当前用途:

  • 稳定人框
  • 为鞋子关联和按人节流提供 track_id

3.5 ai_shoe_det

{
  "id": "shoe_det",
  "type": "ai_shoe_det",
  "role": "filter",
  "enable": true,
  "infer_fps": 2,
  "infer_phase_ms": 150,
  "use_rga": true,
  "use_dma_input": false,
  "model_path": "./models/shoe_detector_openimages_ppe_v1.rknn",
  "model_w": 640,
  "model_h": 640,
  "conf": 0.15,
  "nms": 0.45,
  "v8_box_format": "cxcywh",
  "append_detections": true,
  "dynamic_roi": {
    "enable": true,
    "person_class_id": 0,
    "shoe_class_id": 1,
    "debug_roi_class_id": -1,
    "max_rois": 3,
    "min_person_height": 60,
    "x_offset": -0.24,
    "y_offset": 0.64,
    "width_scale": 1.48,
    "height_scale": 0.58
  }
}

当前用途:

  • 从人框动态生成脚部 ROI
  • 只在 ROI 内跑鞋模型
  • 将鞋框追加回 frame->det

关键参数:

参数 作用 当前建议
conf 鞋候选召回阈值 0.15
append_detections 保留人框并追加鞋框 必须 true
dynamic_roi.max_rois 每帧最多处理多少人 3
dynamic_roi.min_person_height 过滤太远的人 60
x_offset/y_offset/width_scale/height_scale 脚部 ROI 形状 按高机位场景调优

说明:

  • shoe_det.conf 看起来偏低,这是有意为之
  • 当前方案依赖“低阈值召回 + 人鞋关联 + 颜色判断 + 告警节流”整体收敛
  • 在高机位小鞋场景中,如果把这里直接调到 0.40.5,通常会明显漏检

3.6 logic_gate

模式一person_shoe_check

{
  "id": "shoe_assoc",
  "type": "logic_gate",
  "role": "filter",
  "enable": true,
  "mode": "person_shoe_check",
  "person_shoe_check": {
    "person_class": 0,
    "shoe_class": 1,
    "violation_class": 2,
    "min_person_score": 0.30,
    "min_shoe_score": 0.15,
    "require_person_track_id": true,
    "attach_person_track_to_shoe": true,
    "emit_missing_violation": false,
    "foot_region": {
      "x_offset": -0.24,
      "y_offset": 0.64,
      "width_scale": 1.48,
      "height_scale": 0.58
    }
  }
}

当前用途:

  • 把鞋关联到对应的人
  • 把人的 track_id 传给鞋
  • 不再输出“没鞋违规”

模式二ppe_boots_check

{
  "id": "shoe_color",
  "type": "logic_gate",
  "role": "filter",
  "enable": true,
  "mode": "ppe_boots_check",
  "anchor_class": 0,
  "boots_class": 1,
  "violation_class": 2,
  "color_check": {
    "enable": true,
    "method": "brightness",
    "dark_threshold": 90,
    "roi_expand": 1.0
  }
}

当前用途:

  • 只对已检测到的鞋做颜色判断
  • 深色鞋:视为合规
  • 非深色鞋:追加 cls=2 的违规框,供 OSD 和 alarm 使用

颜色参数建议:

参数 含义 当前建议
method 颜色判断方式 brightness
dark_threshold 深色阈值 90
roi_expand 颜色分析区域扩展 1.0

3.7 osd

{
  "id": "osd",
  "type": "osd",
  "role": "filter",
  "enable": true,
  "draw_bbox": true,
  "draw_text": false,
  "use_rga_bbox": false,
  "labels": ["person", "shoe", "non_black_shoe"]
}

当前显示含义:

  • person
  • shoe
  • non_black_shoe

3.8 publish

{
  "id": "pub",
  "type": "publish",
  "role": "filter",
  "enable": true,
  "codec": "h264",
  "fps": 30,
  "bitrate_kbps": 2000,
  "mpp_output_timeout_ms": 50,
  "mpp_packet_wait_ms": 10,
  "use_mpp": true,
  "outputs": [
    {"proto": "rtsp_server", "port": 8555, "path": "/live/cam1"}
  ]
}

当前推荐:

  • 输入端优先 FFmpeg CPU decode
  • 输出端继续使用 MPP 编码

3.9 alarm

{
  "id": "alarm",
  "type": "alarm",
  "role": "sink",
  "enable": true,
  "eval_fps": 2,
  "labels": ["person", "shoe", "non_black_shoe"],
  "rules": [
    {
      "name": "non_compliant_workshoe",
      "class_ids": [2],
      "roi": {"x": 0.0, "y": 0.0, "w": 1.0, "h": 1.0},
      "min_score": 0.30,
      "require_track_id": false,
      "min_duration_ms": 800,
      "min_hits": 2,
      "hit_window_ms": 2000,
      "cooldown_ms": 15000,
      "per_track_cooldown_ms": 0
    }
  ],
  "actions": {
    "log": {
      "enable": true,
      "level": "info",
      "include_detections": true,
      "min_interval_ms": 2000
    }
  }
}

当前告警策略:

  • 只对 cls=2 non_black_shoe 告警
  • 蓝框稳定出现 2 次以上,且持续约 800ms,才触发
  • 触发后进入 15s 冷却,避免反复刷屏

4. 连接关系

{
  "edges": [
    ["in", "pre_rgb"],
    ["pre_rgb", "person_det"],
    ["person_det", "person_trk"],
    ["person_trk", "shoe_det"],
    ["shoe_det", "shoe_assoc"],
    ["shoe_assoc", "shoe_color"],
    ["shoe_color", "osd"],
    ["osd", "post"],
    ["post", "pub"],
    ["pub", "alarm"]
  ]
}

说明:

  • publish -> alarm 是合法链路
  • alarm 会继续读取前面节点保留下来的 frame->det

5. 推荐配置

配置文件 说明
configs/person_shoe_two_stage_workshoe_alarm_v8s_shoe640.json 单路劳保鞋颜色告警主线配置
configs/person_shoe_two_stage_workshoe_alarm_v8s_shoe640_strict.json 单路劳保鞋颜色告警严格版
configs/full_pipeline_1080p.json 人脸检测/识别 + 劳保鞋颜色告警完整流程

6. 调参顺序

建议按下面顺序调,不要同时乱改:

  1. ai_yolo.confbbox_expand 目标:先保证人体框稳定且脚下留出空间
  2. ai_shoe_det.confdynamic_roi 目标:先让鞋框能出来
  3. logic_gate.color_check.dark_threshold 目标:区分黑色劳保鞋和拖鞋/浅色鞋
  4. alarm.min_hits / min_duration_ms / cooldown_ms 目标:把“会报警”收敛成“稳一点再报警”

6.1 实施人员重点参数

实施时优先看下面这些参数,不建议一开始改其它项。

节点 参数 期望效果 调大后的典型结果 调小后的典型结果
input_rtsp use_ffmpeg / use_mpp 解决固定画面卡顿 use_ffmpeg=true 一般更稳 use_mpp=true 某些源可能卡顿
person_det infer_fps 控制人体检测负载 召回更及时,但更吃 NPU 更省资源,但人框更新更慢
person_det bbox_expand.bottom 给脚部 ROI 留空间 更不容易漏脚 人框更紧,鞋 ROI 更容易裁偏
shoe_det conf 控制鞋框召回和误报平衡 误报减少,但漏鞋增多 鞋框更多,但误报也更多
shoe_det.dynamic_roi max_rois 控制每帧最多处理多少人 覆盖更多人,但 FPS 下降 更省资源,但多人时会漏掉部分鞋
shoe_det.dynamic_roi min_person_height 过滤远景小人 远处小人不参与鞋检 更多远景人进入鞋检
shoe_det.dynamic_roi max_box_area_ratio 过滤接近整块脚区的大误框 大蓝框减少 更容易保留整块 ROI 级误报
shoe_assoc min_shoe_score 控制低分鞋框是否参与人鞋关联 低分误框减少 弱鞋框更容易进入后续逻辑
shoe_color dark_threshold 控制黑鞋/非黑鞋边界 更多深灰鞋被判合规 更多鞋会被判非黑鞋
alarm min_duration_ms 控制要稳定多久才报警 更稳,但慢一点 更灵敏,但更容易闪报
alarm cooldown_ms 控制两次告警间隔 减少重复告警 同一事件会更频繁重复报

6.2 推荐调参动作

现场遇到问题时,优先按下面方式处理:

  • 鞋子漏报多:

    • 先降低 shoe_det.conf
    • 再降低 shoe_assoc.min_shoe_score
    • 必要时降低 dynamic_roi.min_person_height
  • 蓝框误报多:

    • 先提高 shoe_det.conf
    • 再提高 shoe_assoc.min_shoe_score
    • 如出现整块脚区级误框,再降低 max_box_area_ratio
  • 黑色劳保鞋被误报:

    • 提高 shoe_color.dark_threshold
  • 非黑鞋不报警:

    • 降低 shoe_color.dark_threshold
    • 检查 alarm.min_score
    • 检查 alarm.min_duration_ms / min_hits
  • 完整流程 FPS 不够:

    • 先降低 face_det.infer_fps
    • 再降低 face_recog.infer_fps
    • 再考虑降低 dynamic_roi.max_rois

7. 常见问题

Q1: 为什么鞋子检测阈值只有 0.15

因为当前场景是高机位、小鞋目标,单纯靠高阈值会先漏掉大量鞋框。系统依赖多级过滤,而不是单级高阈值。

Q2: 为什么有蓝框却不告警?

优先检查:

  • 是否跑的是 person_shoe_two_stage_workshoe_alarm_v8s_shoe640.json
  • alarm.rules[].class_ids 是否包含 2
  • min_score / min_hits / min_duration_ms 是否过严

Q3: 为什么 RTSP 会固定画面卡顿?

如果 VLC 直拉源流不卡,而项目里卡,优先切到:

"use_ffmpeg": true,
"use_mpp": false

这通常说明问题在输入解码兼容性,而不是 AI 链路。


版本v2.0
更新日期2026-03-15

Pose-Enabled Behavior Configuration

This repository now supports two behavior deployment modes.

Mode 1: Baseline Behavior Graph

Use this when you only need tracked detections and rule-based region behavior:

input_rtsp -> preprocess -> ai_yolo -> tracker -> region_event -> action_recog -> event_fusion

Reference config:

  • configs/sample_region_behavior_intrusion.json

Characteristics:

  • No pose model dependency
  • Lowest compute cost
  • fall and fight rely on bbox-only rules

Mode 2: Pose-Enabled Behavior Graph

Use this when you need pose-aware fall and fight:

input_rtsp -> preprocess -> ai_yolo -> tracker -> ai_pose -> pose_assoc -> region_event -> action_recog -> event_fusion

Reference config:

  • configs/sample_region_behavior_full.json

Characteristics:

  • ai_pose writes Frame.pose
  • pose_assoc assigns PoseItem.track_id
  • action_recog can fuse bbox and pose signals per event rule

Optional Integration Rules

  • ai_pose is optional. Do not add it to graphs that only need intrusion or climb.
  • pose_assoc should be placed after ai_pose and after tracked detections are already available on the frame.
  • action_recog must remain runnable when Frame.pose is missing or empty.

Degradation Semantics

  • Without ai_pose, the graph must still behave correctly with bbox-only behavior logic.
  • With ai_pose but without pose_assoc, pose.track_id is not guaranteed and downstream logic should treat it as optional compatibility mode.
  • If pose inference returns no items for a frame, the rest of the graph must keep using Frame.det and track_id.

Structured action_recog Rules

action_recog supports a structured event format with bbox, pose, and fusion sections.

Example:

{
  "id": "action_evt",
  "type": "action_recog",
  "events": [
    {
      "type": "fall",
      "window_ms": 1500,
      "activate_duration_ms": 300,
      "bbox": {
        "enabled": true,
        "min_drop_pixels": 120,
        "min_aspect_ratio_delta": 0.35
      },
      "pose": {
        "enabled": true,
        "min_torso_drop_pixels": 120,
        "max_upright_ratio": 0.60
      },
      "fusion": {
        "match_mode": "any"
      }
    },
    {
      "type": "fight",
      "window_ms": 1200,
      "activate_duration_ms": 200,
      "bbox": {
        "enabled": true,
        "proximity_pixels": 220,
        "min_motion_pixels": 90
      },
      "pose": {
        "enabled": true,
        "min_wrist_motion_pixels": 120,
        "max_wrist_distance_pixels": 120
      },
      "fusion": {
        "match_mode": "any"
      }
    }
  ]
}

Rules:

  • bbox.enabled=false disables bbox signal evaluation for that event.
  • pose.enabled=false disables pose signal evaluation for that event.
  • fusion.match_mode="any" means either bbox or pose may activate the event.
  • fusion.match_mode="all" means bbox and pose must both activate the event.

Backward compatibility:

  • Old flat keys such as min_drop_pixels and pose_min_torso_drop_pixels are still accepted.
  • New configs should prefer the structured form above.