coder/vision-record
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: 三方案改进检测准确度 - YOLO优先、参数调整、连续性判断

Browse Source
This commit is contained in:
hubian
2026-04-16 22:35:51 +08:00
parent d6693f9fd0
commit 7875cca69e
6 changed files with 272 additions and 105 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
config.py
View File

@@ -29,9 +29,14 @@ DEFAULT_CONFIG = {
"display_limit": 20, # 显示最近多少条
# 检测算法开关
"use_haar_cascade": True, # Haar Cascade 人体检测
"use_mediapipe_face": True, # MediaPipe 人脸检测
"use_face_recognition": True, # face_recognition 人脸识别
"use_yolo": True, # YOLO 检测（最准确）
"use_haar_cascade": False, # Haar Cascade 人体检测（备用）
"use_mediapipe_face": True, # MediaPipe 人脸检测
"use_face_recognition": True, # face_recognition 人脸识别
# 连续性判断配置
"confirm_frames": 3, # 连续几帧确认
"min_detection_confidence": 0.3, # 检测置信度阈值
# AI大模型分析开关
"use_vision_api": False, # 是否使用大模型分析（默认关闭）

77
local_analyzer.py
View File

@@ -134,55 +134,70 @@ class LocalAnalyzer:
# 方法1：MediaPipe 人脸检测 + 人员识别（优先）
if HAS_PERSON_MANAGER and use_mediapipe:
print(f"[LocalAnalyzer] Using MediaPipe face detection...")
print(f"[LocalAnalyzer] Using PersonManager for detection...")
person_result = person_manager.analyze_image(image_path, save_new_person=True)
metrics['person_count'] = person_result['total_count']
metrics['new_persons'] = person_result['new_count']
metrics['known_persons'] = person_result['known_count']
metrics['detection_source'] = person_result.get('detection_source', 'unknown')
prev_person_count = self.prev_human_count
person_count_change = person_result['total_count'] - prev_person_count
metrics['person_count_change'] = person_count_change
current_count = person_result['current_count']
person_count_change = current_count - prev_person_count
# 记录人员事件
for person in person_result['persons']:
if person['is_new']:
# 只有确认的变化才记录
if person_result['confirmed_change']:
metrics['person_count_change'] = person_count_change
# 记录人员事件
for person in person_result['persons']:
if person['is_new']:
events.append({
'event_type': '人物活动',
'description': f'新人出现: {person["name"]}，当前共 {current_count} 人',
'confidence': '高',
'source': 'local'
})
self.human_count += 1
self.person_change_count += 1
else:
events.append({
'event_type': '人物活动',
'description': f'已知人员: {person["name"]} [{person.get("source", "detected")}]',
'confidence': '高',
'source': 'local'
})
# 检测人员进出
if person_count_change > 0:
events.append({
'event_type': '人物活动',
'description': f'新人出现: {person["name"]}，当前共 {person_result["total_count"]} 人',
'event_type': '人员进出',
'description': f'检测到 {person_count_change} 人进入，当前共 {current_count} 人 [{person_result.get("detection_source", "")}]',
'confidence': '高',
'source': 'local'
})
self.human_count += 1
self.person_change_count += 1
else:
elif person_count_change < 0:
events.append({
'event_type': '人物活动',
'description': f'已知人员: {person["name"]}',
'event_type': '人员进出',
'description': f'检测到 {abs(person_count_change)} 人离开，当前剩 {current_count} 人',
'confidence': '高',
'source': 'local'
})
self.person_change_count += 1
# 检测人员进出
if person_count_change > 0:
events.append({
'event_type': '人员进出',
'description': f'检测到 {person_count_change} 人进入，当前共 {person_result["total_count"]} 人',
'confidence': '高',
'source': 'local'
})
self.person_change_count += 1
elif person_count_change < 0:
events.append({
'event_type': '人员进出',
'description': f'检测到 {abs(person_count_change)} 人离开，当前剩 {person_result["total_count"]} 人',
'confidence': '高',
'source': 'local'
})
self.person_change_count += 1
self.prev_human_count = person_result['total_count']
self.prev_human_count = current_count
else:
# 没有确认的变化，只记录当前状态
metrics['person_count_change'] = 0
if current_count > 0:
events.append({
'event_type': '人物活动',
'description': f'检测到 {current_count} 人（状态稳定）',
'confidence': '低',
'source': 'local'
})
# 方法2：Haar Cascade 人体检测（备用或并行）
if use_haar and self.human_cascade is not None:

232
person_manager.py
View File

@@ -42,12 +42,20 @@ class PersonManager:
# 加载人员库
self.persons = self._load_persons_db()
# 检测器状态
# 初始化检测器状态
self.face_detector = None
self.mp_face_detection = None
self.cv_face_detector = None
self.has_mediapipe = HAS_MEDIAPIPE
# 从配置读取参数
try:
from config import config_mgr
self.config['mediapipe_min_confidence'] = config_mgr.get('min_detection_confidence', 0.3)
self.config['confirm_frames'] = config_mgr.get('confirm_frames', 3)
except:
pass
# 初始化检测器
self._init_detectors()
@@ -56,8 +64,23 @@ class PersonManager:
'face_match_threshold': 0.6, # 人脸匹配阈值
'unknown_person_id': 'unknown', # 未知人员ID
'max_persons': 100, # 最大人员数量
# 方案1: 参数调整
'mediapipe_min_confidence': 0.3, # 降低阈值，更容易检测
'mediapipe_model_selection': 1, # 1: 远距离模型
'haar_scale_factor': 1.05, # Haar更细粒度
'haar_min_neighbors': 2, # 降低邻居要求
# 方案2: 连续性判断
'confirm_frames': 3, # 连续几帧确认
'leave_frames': 2, # 连续几帧消失才算离开
}
# 方案2: 追踪状态（连续判断）
self.tracked_persons = {} # {person_id: {'frames': count, 'confirmed': bool}}
self.prev_persons = [] # 前一帧检测到的人
self.confirmation_buffer = {} # 确认缓冲区
# 统计
self.total_detections = 0
self.known_persons_detected = 0
@@ -80,17 +103,16 @@ class PersonManager:
def _init_detectors(self):
"""初始化检测器"""
# MediaPipe 人脸检测
# MediaPipe 人脸检测（方案1: 参数调整）
if self.has_mediapipe:
try:
# 使用更安全的导入方式
mp_face_detection = mp.solutions.face_detection
self.face_detector = mp_face_detection.FaceDetection(
model_selection=0, # 0: 短距离，1: 远距离
min_detection_confidence=0.5
model_selection=self.config['mediapipe_model_selection'], # 远距离模型
min_detection_confidence=self.config['mediapipe_min_confidence'] # 降低阈值
)
self.mp_face_detection = mp_face_detection
print("[PersonManager] MediaPipe face detector initialized")
print(f"[PersonManager] MediaPipe initialized (model={self.config['mediapipe_model_selection']}, conf={self.config['mediapipe_min_confidence']})")
except Exception as e:
print(f"[PersonManager] MediaPipe init failed: {e}")
self.face_detector = None
@@ -101,13 +123,25 @@ class PersonManager:
model_path = cv2.data.haarcascades + 'haarcascade_frontalface_default.xml'
if Path(model_path).exists():
self.cv_face_detector = cv2.CascadeClassifier(model_path)
print("[PersonManager] OpenCV face detector initialized (backup)")
print("[PersonManager] OpenCV Haar Cascade initialized (backup)")
except Exception as e:
self.cv_face_detector = None
print(f"[PersonManager] OpenCV face detector init failed: {e}")
print(f"[PersonManager] OpenCV detector init failed: {e}")
# 方案3: YOLO 检测（更准确）
self.yolo_detector = None
try:
from ultralytics import YOLO
# 使用轻量级 nano 模型
self.yolo_detector = YOLO('yolov8n.pt') # nano 模型，快速
print("[PersonManager] YOLOv8nano initialized (most accurate)")
except ImportError:
print("[PersonManager] YOLO not installed. Install with: pip install ultralytics")
except Exception as e:
print(f"[PersonManager] YOLO init failed: {e}")
def detect_faces(self, image):
"""检测人脸
"""检测人脸（优先使用 YOLO，其次 MediaPipe，最后 Haar）
Args:
image: 图片（numpy array 或路径）
@@ -123,7 +157,31 @@ class PersonManager:
faces = []
# MediaPipe 检测
# 方案3: YOLO 检测（优先，最准确）
if self.yolo_detector is not None:
try:
results = self.yolo_detector(image, classes=[0], verbose=False) # class 0 = person
for r in results:
for box in r.boxes:
x1, y1, x2, y2 = box.xyxy[0].tolist()
conf = box.conf[0].item()
# 转换为 [x, y, w, h] 格式
faces.append({
'bbox': [int(x1), int(y1), int(x2-x1), int(y2-y1)],
'confidence': conf,
'source': 'yolo'
})
if faces:
print(f"[PersonManager] YOLO detected {len(faces)} persons")
return faces # YOLO 检测成功，直接返回
except Exception as e:
print(f"[PersonManager] YOLO detection failed: {e}")
# 方案1+2: MediaPipe 检测
if self.has_mediapipe and self.face_detector is not None:
rgb_image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
results = self.face_detector.process(rgb_image)
@@ -144,13 +202,16 @@ class PersonManager:
'source': 'mediapipe'
})
# OpenCV 检测（备用）
elif self.cv_face_detector is not None:
if faces:
return faces
# 备用: OpenCV Haar 检测
if self.cv_face_detector is not None:
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
detections = self.cv_face_detector.detectMultiScale(
gray,
scaleFactor=1.1,
minNeighbors=5,
scaleFactor=self.config['haar_scale_factor'],
minNeighbors=self.config['haar_min_neighbors'],
minSize=(30, 30)
)
@@ -351,7 +412,7 @@ class PersonManager:
self._save_persons_db()
def analyze_image(self, image_path, save_new_person=True):
"""分析图片中的人员
"""分析图片中的人员（带连续性判断）
Args:
image_path: 图片路径
@@ -363,6 +424,7 @@ class PersonManager:
'persons': list, # 识别的人员
'new_count': int, # 新人员数量
'known_count': int, # 已知人员数量
'confirmed_change': bool, # 是否有确认的人员变化
}
"""
image = cv2.imread(image_path)
@@ -373,61 +435,117 @@ class PersonManager:
# 检测人脸
faces = self.detect_faces(image)
current_count = len(faces)
persons = []
new_count = 0
known_count = 0
# 方案2: 连续性判断
confirmed_change = False
confirmed_persons = []
for face in faces:
bbox = face['bbox']
# 检查人数变化
prev_count = len(self.prev_persons)
# 提取特征
encoding = self.extract_face_encoding(image, bbox)
if current_count != prev_count:
# 人数变化，记录到缓冲区
key = f"count_{current_count}"
if key not in self.confirmation_buffer:
self.confirmation_buffer[key] = {'count': 0, 'persons': []}
# 匹配
match_result = self.match_face(encoding)
self.confirmation_buffer[key]['count'] += 1
if match_result['is_new']:
# 新人员
new_count += 1
# 临时识别人员
temp_persons = []
for face in faces:
bbox = face['bbox']
encoding = self.extract_face_encoding(image, bbox)
match_result = self.match_face(encoding)
if save_new_person and len(self.persons) < self.config['max_persons']:
new_person = self.add_new_person(image, bbox)
if new_person:
persons.append({
'person_id': new_person['person_id'],
'name': new_person['name'],
'bbox': bbox,
'is_new': True,
'confidence': face['confidence']
})
else:
persons.append({
'person_id': 'unknown',
'name': 'Unknown (new)',
'bbox': bbox,
'is_new': True,
'confidence': face['confidence']
})
else:
# 已知人员
known_count += 1
self.update_person_visit(match_result['person_id'])
persons.append({
'person_id': match_result['person_id'],
person_info = {
'person_id': match_result['person_id'] if not match_result['is_new'] else 'unknown',
'name': match_result['name'],
'bbox': bbox,
'is_new': False,
'confidence': match_result['confidence']
})
'is_new': match_result['is_new'],
'confidence': face['confidence'],
'source': face['source']
}
temp_persons.append(person_info)
self.confirmation_buffer[key]['persons'] = temp_persons
# 达到确认帧数
if self.confirmation_buffer[key]['count'] >= self.config['confirm_frames']:
confirmed_change = True
confirmed_persons = temp_persons
print(f"[PersonManager] Confirmed: {prev_count} -> {current_count} persons (after {self.config['confirm_frames']} frames)")
# 清空其他缓冲区
self.confirmation_buffer = {}
# 更新前一帧状态
self.prev_persons = temp_persons
else:
# 人数不变，清空变化缓冲区，维持当前状态
if current_count > 0:
# 识别当前人员
temp_persons = []
for face in faces:
bbox = face['bbox']
encoding = self.extract_face_encoding(image, bbox)
match_result = self.match_face(encoding)
person_info = {
'person_id': match_result['person_id'] if not match_result['is_new'] else 'unknown',
'name': match_result['name'],
'bbox': bbox,
'is_new': match_result['is_new'],
'confidence': face['confidence'],
'source': face['source']
}
temp_persons.append(person_info)
confirmed_persons = temp_persons
self.prev_persons = temp_persons
# 清空变化缓冲区
keys_to_remove = [k for k in self.confirmation_buffer.keys() if not k.endswith(f"_{current_count}")]
for k in keys_to_remove:
del self.confirmation_buffer[k]
# 统计新人和已知人员
new_count = 0
known_count = 0
persons_to_save = []
for person in confirmed_persons:
if person['is_new']:
new_count += 1
# 只有确认后才保存新人
if confirmed_change and save_new_person and len(self.persons) < self.config['max_persons']:
# 找到对应的 face bbox
for face in faces:
if face['bbox'] == person['bbox']:
new_person = self.add_new_person(image, face['bbox'])
if new_person:
person['person_id'] = new_person['person_id']
person['name'] = new_person['name']
persons_to_save.append(person)
break
else:
known_count += 1
self.update_person_visit(person['person_id'])
persons_to_save.append(person)
return {
'faces': faces,
'persons': persons,
'persons': persons_to_save,
'new_count': new_count,
'known_count': known_count,
'total_count': len(persons)
'total_count': len(persons_to_save),
'confirmed_change': confirmed_change,
'current_count': current_count,
'prev_count': prev_count,
'detection_source': faces[0]['source'] if faces else 'none'
}
def get_persons_list(self):

6
requirements.txt
View File

@@ -4,6 +4,6 @@ uvicorn>=0.23.0
requests>=2.31.0
numpy>=1.20.0
# Optional: More accurate face detection and recognition
# mediapipe>=0.10.0
# face-recognition>=1.7.0 (requires dlib, may need manual install on Windows)
# Optional: More accurate detection
ultralytics>=8.0.0
mediapipe>=0.10.0

12
web/static/app.js
View File

@@ -507,10 +507,15 @@ function loadSettingsForm() {
document.getElementById('setting-refresh-interval').value = config.refresh_interval || 5;
// Detection algorithm settings
document.getElementById('setting-use-haar').checked = config.use_haar_cascade !== false;
document.getElementById('setting-use-yolo').checked = config.use_yolo !== false;
document.getElementById('setting-use-mediapipe').checked = config.use_mediapipe_face !== false;
document.getElementById('setting-use-haar').checked = config.use_haar_cascade === true;
document.getElementById('setting-use-face-rec').checked = config.use_face_recognition !== false;
// Confirmation settings
document.getElementById('setting-confirm-frames').value = config.confirm_frames || 3;
document.getElementById('setting-min-confidence').value = config.min_detection_confidence || 0.3;
// Vision API settings
document.getElementById('setting-use-vision-api').checked = config.use_vision_api === true;
document.getElementById('setting-vision-trigger').value = config.vision_api_trigger || 'person_change';
@@ -533,10 +538,15 @@ function saveSettings() {
refresh_interval: parseInt(document.getElementById('setting-refresh-interval').value),
// Detection algorithms
use_yolo: document.getElementById('setting-use-yolo').checked,
use_haar_cascade: document.getElementById('setting-use-haar').checked,
use_mediapipe_face: document.getElementById('setting-use-mediapipe').checked,
use_face_recognition: document.getElementById('setting-use-face-rec').checked,
// Confirmation settings
confirm_frames: parseInt(document.getElementById('setting-confirm-frames').value),
min_detection_confidence: parseFloat(document.getElementById('setting-min-confidence').value),
// Vision API
use_vision_api: document.getElementById('setting-use-vision-api').checked,
vision_api_trigger: document.getElementById('setting-vision-trigger').value,

35
web/static/index.html
View File

@@ -138,26 +138,45 @@
</div>
<div class="settings-section">
<h4>检测算法设置</h4>
<h4>Detection Algorithms</h4>
<div class="setting-item">
<label>Haar Cascade 人体检测:</label>
<input type="checkbox" id="setting-use-haar" checked>
<span class="setting-desc">传统人体检测（备用）</span>
<label>YOLO (Most Accurate):</label>
<input type="checkbox" id="setting-use-yolo" checked>
<span class="setting-desc">YOLOv8 nano - Best accuracy</span>
</div>
<div class="setting-item">
<label>MediaPipe 人脸检测:</label>
<label>MediaPipe Face:</label>
<input type="checkbox" id="setting-use-mediapipe" checked>
<span class="setting-desc">高精度人脸检测</span>
<span class="setting-desc">High precision face detection</span>
</div>
<div class="setting-item">
<label>Haar Cascade Body:</label>
<input type="checkbox" id="setting-use-haar">
<span class="setting-desc">Traditional body detection (backup)</span>
</div>
<div class="setting-item">
<label>Face Recognition:</label>
<input type="checkbox" id="setting-use-face-rec" checked>
<span class="setting-desc">人脸识别（识别同一人）</span>
<span class="setting-desc">Identify same person</span>
</div>
</div>
<div class="settings-section">
<h4>AI大模型分析</h4>
<h4>Confirmation Settings</h4>
<div class="setting-item">
<label>Confirm Frames:</label>
<input type="number" id="setting-confirm-frames" value="3" min="1" max="10">
<span class="setting-desc">Frames to confirm detection</span>
</div>
<div class="setting-item">
<label>Min Confidence:</label>
<input type="number" id="setting-min-confidence" value="0.3" min="0.1" max="1" step="0.1">
<span class="setting-desc">Detection confidence threshold</span>
</div>
</div>
<div class="settings-section">
<h4>AI Analysis</h4>
<div class="setting-item">
<label>启用大模型分析:</label>
<input type="checkbox" id="setting-use-vision-api">