馴龍窩

加上可以用滑鼠左鍵拉偵測ROI範圍執行
python ai_face_stream_native-v4.py

import os
import sys
import time
import numpy as np
import cv2
import threading
import queue
# --- 0. 語音功能設定 (TTS) ---
speech_queue = queue.Queue()
is_face_present = False # 記錄目前是否有人臉的狀態 (狀態機)
last_face_time = 0 # 記錄最後一次看到人臉的時間
FACE_TIMEOUT = 3.0 # 離開畫面超過 3 秒才重置狀態
def tts_worker():
"""獨立的語音執行緒，避免阻塞影像串流"""
if os.name == 'nt':
# Windows 環境：直接使用原生的 SAPI5，避開 pyttsx3 的卡死 Bug
try:
import pythoncom
import win32com.client
# 必須在 Thread 中初始化 COM
pythoncom.CoInitialize()
speaker = win32com.client.Dispatch("SAPI.SpVoice")
print("✅ 成功載入 Windows 原生語音引擎 (win32com)")
while True:
text = speech_queue.get()
if text is None:
break
print(f"🗣️ [語音播報] {text}")
speaker.Speak(text)
return
except Exception as e:
print(f"⚠️ win32com 語音初始化失敗: {e}，將嘗試備用方案...")
# 備用方案：pyttsx3
try:
import pyttsx3
except ImportError:
print("❌ 找不到語音模組，語音功能停用。")
return
while True:
text = speech_queue.get()
if text is None:
break
print(f"🗣️ [語音播報] {text}")
try:
# 每次發聲前重新 init，發聲完後銷毀，避免 runAndWait 卡死
engine = pyttsx3.init()
engine.say(text)
engine.runAndWait()
except Exception as e:
print(f"TTS 錯誤: {e}")
# 啟動語音執行緒
tts_thread = threading.Thread(target=tts_worker, daemon=True)
tts_thread.start()
# --- 1. Windows 環境設定 ---
if os.name == 'nt':
gst_root = r"C:\Program Files\gstreamer\1.0\msvc_x86_64"
path_bin = os.path.join(gst_root, "bin")
path_plugins = os.path.join(gst_root, "lib", "gstreamer-1.0")
if os.path.exists(path_bin):
os.add_dll_directory(path_bin)
os.environ['PATH'] = path_bin + ";" + os.environ['PATH']
if os.path.exists(path_plugins):
os.environ['GST_PLUGIN_PATH'] = path_plugins
try:
import gi
gi.require_version('Gst', '1.0')
from gi.repository import Gst, GLib
except ImportError:
print("❌ GStreamer 載入失敗")
sys.exit(1)
Gst.init(None)
# --- 2. 設定接收端 IP ---
RECEIVER_IP = "127.0.0.1"
RECEIVER_PORT = 5000
# --- 3. 建立發送管線 (Sender Pipeline) ---
sender_cmd = f"""
appsrc name=mysource format=3 is-live=True !
video/x-raw,format=BGR,width=1280,height=720,framerate=30/1 !
videoconvert !
openh264enc complexity=0 bitrate=3000000 !
rtph264pay config-interval=1 pt=96 !
udpsink host={RECEIVER_IP} port={RECEIVER_PORT} sync=false
"""
print("建立發送管線...")
try:
sender_pipeline = Gst.parse_launch(sender_cmd)
sender_src = sender_pipeline.get_by_name("mysource")
sender_pipeline.set_state(Gst.State.PLAYING)
print(f"📡 串流發射準備就緒 -> {RECEIVER_IP}:{RECEIVER_PORT}")
except Exception as e:
print(f"❌ 發送管線建立失敗: {e}")
sys.exit(1)
# --- 4. 載入 AI 模型 ---
protoPath = "deploy.prototxt"
modelPath = "res10_300x300_ssd_iter_140000.caffemodel"
if not os.path.exists(protoPath) or not os.path.exists(modelPath):
print("❌ 找不到模型檔案")
sys.exit(1)
net = cv2.dnn.readNetFromCaffe(protoPath, modelPath)
print("✅ AI 模型載入成功！")
prev_frame_time = 0
# --- 定義滑鼠拖曳 ROI 的變數與回呼函數 ---
drawing = False # 記錄是否正在拖曳滑鼠
ix, iy = -1, -1
# 初始預設的偵測區域
ROI_X1, ROI_Y1 = 320, 120
ROI_X2, ROI_Y2 = 960, 600
def draw_roi(event, x, y, flags, param):
"""處理滑鼠事件，供使用者自定義 ROI 範圍"""
global ix, iy, drawing, ROI_X1, ROI_Y1, ROI_X2, ROI_Y2
# 由於在畫面上顯示的 img (small_frame) 被縮小為 0.5 倍
# 這裡必須將滑鼠坐標 x, y 乘以 2 才能對應回原始 1280x720 的影像座標
real_x = x * 2
real_y = y * 2
if event == cv2.EVENT_LBUTTONDOWN:
drawing = True
ix, iy = real_x, real_y
ROI_X1, ROI_Y1 = ix, iy
ROI_X2, ROI_Y2 = ix, iy
elif event == cv2.EVENT_MOUSEMOVE:
if drawing:
ROI_X2, ROI_Y2 = real_x, real_y
elif event == cv2.EVENT_LBUTTONUP:
drawing = False
ROI_X2, ROI_Y2 = real_x, real_y
window_initialized = False # 紀錄控制面板是否已初始化
UI_WINDOW_NAME = 'AI Face Stream' # 統一的 UI 視窗名稱
# --- 5. 核心處理函數 ---
def on_new_sample(sink):
global prev_frame_time, is_face_present, last_face_time, window_initialized
global ROI_X1, ROI_Y1, ROI_X2, ROI_Y2, drawing
sample = sink.emit("pull-sample")
buf = sample.get_buffer()
caps = sample.get_caps()
height = caps.get_structure(0).get_value('height')
width = caps.get_structure(0).get_value('width')
result, mapinfo = buf.map(Gst.MapFlags.READ)
if result:
new_frame_time = time.time()
fps = 1 / (new_frame_time - prev_frame_time) if (new_frame_time - prev_frame_time) > 0 else 0
prev_frame_time = new_frame_time
# 轉成圖片
img_rgb = np.ndarray((height, width, 3), buffer=mapinfo.data, dtype=np.uint8)
img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
# --- 初始化整合介面 (只執行一次) ---
if not window_initialized:
cv2.namedWindow(UI_WINDOW_NAME, cv2.WINDOW_NORMAL)
# 留一點高度空間給滑桿
cv2.resizeWindow(UI_WINDOW_NAME, 800, 700)
# Zoom: 10~30 代表 1.0x ~ 3.0x
cv2.createTrackbar('Zoom (x10)', UI_WINDOW_NAME, 10, 30, lambda x: None)
# 銳利度: 0~10 代表 0 ~ 1.0 的權重
cv2.createTrackbar('Sharpness', UI_WINDOW_NAME, 0, 10, lambda x: None)
# 飽和度: 0~200 代表 0.0x ~ 2.0x (100 為正常)
cv2.createTrackbar('Saturation %', UI_WINDOW_NAME, 100, 200, lambda x: None)
# RGB 調整: 0~200 代表 0.0x ~ 2.0x (100 為正常)
cv2.createTrackbar('R %', UI_WINDOW_NAME, 100, 200, lambda x: None)
cv2.createTrackbar('G %', UI_WINDOW_NAME, 100, 200, lambda x: None)
cv2.createTrackbar('B %', UI_WINDOW_NAME, 100, 200, lambda x: None)
# 註冊滑鼠回呼事件
cv2.setMouseCallback(UI_WINDOW_NAME, draw_roi)
window_initialized = True
# --- 讀取控制面板數值 ---
try:
zoom_val = max(1.0, cv2.getTrackbarPos('Zoom (x10)', UI_WINDOW_NAME) / 10.0)
sharp_val = cv2.getTrackbarPos('Sharpness', UI_WINDOW_NAME) / 10.0
sat_val = cv2.getTrackbarPos('Saturation %', UI_WINDOW_NAME) / 100.0
r_val = cv2.getTrackbarPos('R %', UI_WINDOW_NAME) / 100.0
g_val = cv2.getTrackbarPos('G %', UI_WINDOW_NAME) / 100.0
b_val = cv2.getTrackbarPos('B %', UI_WINDOW_NAME) / 100.0
except cv2.error:
# 避免視窗尚未完全就緒時讀取報錯
zoom_val, sharp_val, sat_val = 1.0, 0.0, 1.0
r_val, g_val, b_val = 1.0, 1.0, 1.0
# --- 影像處理 ---
# 1. Zoom (縮放：裁切畫面正中央後放大回原尺寸)
if zoom_val > 1.0:
new_w, new_h = int(width / zoom_val), int(height / zoom_val)
left = (width - new_w) // 2
top = (height - new_h) // 2
cropped = img_bgr[top:top+new_h, left:left+new_w]
img_bgr = cv2.resize(cropped, (width, height))
# 2. RGB 調整 (乘上對應的比例)
if r_val != 1.0 or g_val != 1.0 or b_val != 1.0:
img_bgr = np.clip(img_bgr * np.array([b_val, g_val, r_val]), 0, 255).astype(np.uint8)
# 3. Saturation (飽和度)
if sat_val != 1.0:
hsv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2HSV).astype(np.float32)
hsv[:, :, 1] *= sat_val
hsv[:, :, 1] = np.clip(hsv[:, :, 1], 0, 255)
img_bgr = cv2.cvtColor(hsv.astype(np.uint8), cv2.COLOR_HSV2BGR)
# 4. Sharpness (銳利度)
if sharp_val > 0:
kernel = np.array([[-1, -1, -1], [-1, 9, -1], [-1, -1, -1]])
sharpened = cv2.filter2D(img_bgr, -1, kernel)
img_bgr = cv2.addWeighted(img_bgr, 1.0 - sharp_val, sharpened, sharp_val, 0)
# --- 處理與檢查 ROI 邊界防呆 ---
# 確保 x1 一定小於 x2，以應對由右下往左上反向拖曳的狀況
cur_x1, cur_x2 = min(ROI_X1, ROI_X2), max(ROI_X1, ROI_X2)
cur_y1, cur_y2 = min(ROI_Y1, ROI_Y2), max(ROI_Y1, ROI_Y2)
# 限制範圍不能超出實際影像大小
cur_x1 = max(0, min(cur_x1, width - 1))
cur_y1 = max(0, min(cur_y1, height - 1))
cur_x2 = max(0, min(cur_x2, width - 1))
cur_y2 = max(0, min(cur_y2, height - 1))
# 避免畫出的框太小或變為單點導致程式或 AI 崩潰 (最少 10x10)
if cur_x2 - cur_x1 < 10:
cur_x2 = cur_x1 + 10 if cur_x1 + 10 < width else cur_x2
if cur_y2 - cur_y1 < 10:
cur_y2 = cur_y1 + 10 if cur_y1 + 10 < height else cur_y2
roi_w = cur_x2 - cur_x1
roi_h = cur_y2 - cur_y1
# 裁切出特定區域 (ROI)
roi_img = img_bgr[cur_y1:cur_y2, cur_x1:cur_x2]
# --- AI 偵測 (針對 ROI) ---
blob = cv2.dnn.blobFromImage(cv2.resize(roi_img, (300, 300)), 1.0, (300, 300), (104.0, 177.0, 123.0))
net.setInput(blob)
detections = net.forward()
face_count = 0
# 畫出偵測區域的框線
# 如果使用者正在拖曳中，框線換成橘黃色，放開後變藍色
roi_color = (0, 165, 255) if drawing else (255, 0, 0)
cv2.rectangle(img_bgr, (cur_x1, cur_y1), (cur_x2, cur_y2), roi_color, 2)
cv2.putText(img_bgr, "Detection Area", (cur_x1, cur_y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.6, roi_color, 2)
for i in range(0, detections.shape[2]):
confidence = detections[0, 0, i, 2]
if confidence > 0.6:
face_count += 1
# 計算在 ROI 內的座標
box = detections[0, 0, i, 3:7] * np.array([roi_w, roi_h, roi_w, roi_h])
(startX, startY, endX, endY) = box.astype("int")
# 將 ROI 的座標轉換回「原始全畫面」的座標，才能畫在正確位置
startX += cur_x1
startY += cur_y1
endX += cur_x1
endY += cur_y1
text = f"{confidence * 100:.2f}%"
y = startY - 10 if startY - 10 > 10 else startY + 10
cv2.rectangle(img_bgr, (startX, startY), (endX, endY), (0, 255, 0), 2)
cv2.putText(img_bgr, text, (startX, y), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
current_time = time.time()
# --- 判斷與顯示狀態 (狀態機防閃爍機制) ---
if face_count > 0:
status_text = f"Status: Face Detected (Count: {face_count})"
status_color = (0, 255, 0) # 綠色
# 狀態機：如果原本沒有人臉，現在偵測到了 -> 觸發語音
if not is_face_present:
is_face_present = True
print(">>> 🟢 狀態切換：進入 ROI，偵測到人臉！")
speech_queue.put("偵測到人臉")
# 更新最後看到人臉的時間
last_face_time = current_time
else:
status_text = "Status: No Face"
status_color = (0, 0, 255) # 紅色
# 狀態機：如果原本有人臉，且離開畫面已經超過 FACE_TIMEOUT 秒 -> 重置狀態
if is_face_present and (current_time - last_face_time > FACE_TIMEOUT):
is_face_present = False
print(f">>> 🔴 狀態切換：人臉已離開 ROI 超過 {FACE_TIMEOUT} 秒，狀態重置！")
# 顯示狀態、人臉個數與操作提示
cv2.putText(img_bgr, status_text, (20, 90), cv2.FONT_HERSHEY_SIMPLEX, 1.0, status_color, 3)
cv2.putText(img_bgr, f"FPS: {int(fps)}", (20, 50), cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0, 255, 255), 3)
cv2.putText(img_bgr, "Tip: Drag mouse to draw Detection Area", (20, 130), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (200, 200, 200), 2)
# --- 發送串流 ---
data = img_bgr.tobytes()
gst_buffer = Gst.Buffer.new_allocate(None, len(data), None)
gst_buffer.fill(0, data)
gst_buffer.pts = buf.pts
gst_buffer.dts = buf.dts
gst_buffer.duration = buf.duration
sender_src.emit("push-buffer", gst_buffer)
# 本地顯示 (整合在同一個 UI)
# 注意: 我們在這裡將畫面縮小為 0.5，所以滑鼠回呼事件需要乘以 2
small_frame = cv2.resize(img_bgr, None, fx=0.5, fy=0.5)
cv2.imshow(UI_WINDOW_NAME, small_frame)
cv2.waitKey(1)
buf.unmap(mapinfo)
return Gst.FlowReturn.OK
# --- 6. 建立攝影機接收管線 ---
cmd = """
mfvideosrc device-index=0 ! decodebin ! videoconvert ! videoscale ! video/x-raw, width=1280, height=720, format=RGB ! appsink name=mysink emit-signals=True drop=True
"""
pipeline = Gst.parse_launch(cmd)
appsink = pipeline.get_by_name("mysink")
appsink.connect("new-sample", on_new_sample)
print("🚀 系統啟動中... (按 Ctrl+C 停止)")
pipeline.set_state(Gst.State.PLAYING)
loop = GLib.MainLoop()
try:
loop.run()
except KeyboardInterrupt:
print("停止中...")
pipeline.set_state(Gst.State.NULL)
sender_pipeline.set_state(Gst.State.NULL)
speech_queue.put(None) # 關閉語音執行緒
cv2.destroyAllWindows()

繼上次，加上ROI之後，這次繼續加上數位變焦、銳利度、對比度、R、G、B調整功能
執行
python ai_face_stream_native-v3.py
ai_face_stream_native-v3.py 原始碼如下:

import os
import sys
import time
import numpy as np
import cv2
import threading
import queue
# --- 0. 語音功能設定 (TTS) ---
speech_queue = queue.Queue()
is_face_present = False # 記錄目前是否有人臉的狀態 (狀態機)
last_face_time = 0 # 記錄最後一次看到人臉的時間
FACE_TIMEOUT = 3.0 # 離開畫面超過 3 秒才重置狀態
def tts_worker():
"""獨立的語音執行緒，避免阻塞影像串流"""
if os.name == 'nt':
# Windows 環境：直接使用原生的 SAPI5，避開 pyttsx3 的卡死 Bug
try:
import pythoncom
import win32com.client
# 必須在 Thread 中初始化 COM
pythoncom.CoInitialize()
speaker = win32com.client.Dispatch("SAPI.SpVoice")
print("✅ 成功載入 Windows 原生語音引擎 (win32com)")
while True:
text = speech_queue.get()
if text is None:
break
print(f"🗣️ [語音播報] {text}")
speaker.Speak(text)
return
except Exception as e:
print(f"⚠️ win32com 語音初始化失敗: {e}，將嘗試備用方案...")
# 備用方案：pyttsx3
try:
import pyttsx3
except ImportError:
print("❌ 找不到語音模組，語音功能停用。")
return
while True:
text = speech_queue.get()
if text is None:
break
print(f"🗣️ [語音播報] {text}")
try:
# 每次發聲前重新 init，發聲完後銷毀，避免 runAndWait 卡死
engine = pyttsx3.init()
engine.say(text)
engine.runAndWait()
except Exception as e:
print(f"TTS 錯誤: {e}")
# 啟動語音執行緒
tts_thread = threading.Thread(target=tts_worker, daemon=True)
tts_thread.start()
# --- 1. Windows 環境設定 ---
if os.name == 'nt':
gst_root = r"C:\Program Files\gstreamer\1.0\msvc_x86_64"
path_bin = os.path.join(gst_root, "bin")
path_plugins = os.path.join(gst_root, "lib", "gstreamer-1.0")
if os.path.exists(path_bin):
os.add_dll_directory(path_bin)
os.environ['PATH'] = path_bin + ";" + os.environ['PATH']
if os.path.exists(path_plugins):
os.environ['GST_PLUGIN_PATH'] = path_plugins
try:
import gi
gi.require_version('Gst', '1.0')
from gi.repository import Gst, GLib
except ImportError:
print("❌ GStreamer 載入失敗")
sys.exit(1)
Gst.init(None)
# --- 2. 設定接收端 IP ---
RECEIVER_IP = "127.0.0.1"
RECEIVER_PORT = 5000
# --- 3. 建立發送管線 (Sender Pipeline) ---
sender_cmd = f"""
appsrc name=mysource format=3 is-live=True !
video/x-raw,format=BGR,width=1280,height=720,framerate=30/1 !
videoconvert !
openh264enc complexity=0 bitrate=3000000 !
rtph264pay config-interval=1 pt=96 !
udpsink host={RECEIVER_IP} port={RECEIVER_PORT} sync=false
"""
print("建立發送管線...")
try:
sender_pipeline = Gst.parse_launch(sender_cmd)
sender_src = sender_pipeline.get_by_name("mysource")
sender_pipeline.set_state(Gst.State.PLAYING)
print(f"📡 串流發射準備就緒 -> {RECEIVER_IP}:{RECEIVER_PORT}")
except Exception as e:
print(f"❌ 發送管線建立失敗: {e}")
sys.exit(1)
# --- 4. 載入 AI 模型 ---
protoPath = "deploy.prototxt"
modelPath = "res10_300x300_ssd_iter_140000.caffemodel"
if not os.path.exists(protoPath) or not os.path.exists(modelPath):
print("❌ 找不到模型檔案")
sys.exit(1)
net = cv2.dnn.readNetFromCaffe(protoPath, modelPath)
print("✅ AI 模型載入成功！")
prev_frame_time = 0
# --- 定義偵測區域 (Region of Interest) ---
# 假設原始影像是 1280x720，我們在正中間開一個 640x480 的區域
ROI_X1, ROI_Y1 = 320, 120
ROI_X2, ROI_Y2 = 960, 600
ROI_W = ROI_X2 - ROI_X1
ROI_H = ROI_Y2 - ROI_Y1
window_initialized = False # 紀錄控制面板是否已初始化
UI_WINDOW_NAME = 'AI Face Stream' # 統一的 UI 視窗名稱
# --- 5. 核心處理函數 ---
def on_new_sample(sink):
global prev_frame_time, is_face_present, last_face_time, window_initialized
sample = sink.emit("pull-sample")
buf = sample.get_buffer()
caps = sample.get_caps()
height = caps.get_structure(0).get_value('height')
width = caps.get_structure(0).get_value('width')
result, mapinfo = buf.map(Gst.MapFlags.READ)
if result:
new_frame_time = time.time()
fps = 1 / (new_frame_time - prev_frame_time) if (new_frame_time - prev_frame_time) > 0 else 0
prev_frame_time = new_frame_time
# 轉成圖片
img_rgb = np.ndarray((height, width, 3), buffer=mapinfo.data, dtype=np.uint8)
img_bgr = cv2.cvtColor(img_rgb, cv2.COLOR_RGB2BGR)
# --- 初始化整合介面 (只執行一次) ---
if not window_initialized:
cv2.namedWindow(UI_WINDOW_NAME, cv2.WINDOW_NORMAL)
# 留一點高度空間給滑桿
cv2.resizeWindow(UI_WINDOW_NAME, 800, 700)
# Zoom: 10~30 代表 1.0x ~ 3.0x
cv2.createTrackbar('Zoom (x10)', UI_WINDOW_NAME, 10, 30, lambda x: None)
# 銳利度: 0~10 代表 0 ~ 1.0 的權重
cv2.createTrackbar('Sharpness', UI_WINDOW_NAME, 0, 10, lambda x: None)
# 飽和度: 0~200 代表 0.0x ~ 2.0x (100 為正常)
cv2.createTrackbar('Saturation %', UI_WINDOW_NAME, 100, 200, lambda x: None)
# RGB 調整: 0~200 代表 0.0x ~ 2.0x (100 為正常)
cv2.createTrackbar('R %', UI_WINDOW_NAME, 100, 200, lambda x: None)
cv2.createTrackbar('G %', UI_WINDOW_NAME, 100, 200, lambda x: None)
cv2.createTrackbar('B %', UI_WINDOW_NAME, 100, 200, lambda x: None)
window_initialized = True
# --- 讀取控制面板數值 ---
try:
zoom_val = max(1.0, cv2.getTrackbarPos('Zoom (x10)', UI_WINDOW_NAME) / 10.0)
sharp_val = cv2.getTrackbarPos('Sharpness', UI_WINDOW_NAME) / 10.0
sat_val = cv2.getTrackbarPos('Saturation %', UI_WINDOW_NAME) / 100.0
r_val = cv2.getTrackbarPos('R %', UI_WINDOW_NAME) / 100.0
g_val = cv2.getTrackbarPos('G %', UI_WINDOW_NAME) / 100.0
b_val = cv2.getTrackbarPos('B %', UI_WINDOW_NAME) / 100.0
except cv2.error:
# 避免視窗尚未完全就緒時讀取報錯
zoom_val, sharp_val, sat_val = 1.0, 0.0, 1.0
r_val, g_val, b_val = 1.0, 1.0, 1.0
# --- 影像處理 ---
# 1. Zoom (縮放：裁切畫面正中央後放大回原尺寸)
if zoom_val > 1.0:
new_w, new_h = int(width / zoom_val), int(height / zoom_val)
left = (width - new_w) // 2
top = (height - new_h) // 2
cropped = img_bgr[top:top+new_h, left:left+new_w]
img_bgr = cv2.resize(cropped, (width, height))
# 2. RGB 調整 (乘上對應的比例)
if r_val != 1.0 or g_val != 1.0 or b_val != 1.0:
img_bgr = np.clip(img_bgr * np.array([b_val, g_val, r_val]), 0, 255).astype(np.uint8)
# 3. Saturation (飽和度)
if sat_val != 1.0:
hsv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2HSV).astype(np.float32)
hsv[:, :, 1] *= sat_val
hsv[:, :, 1] = np.clip(hsv[:, :, 1], 0, 255)
img_bgr = cv2.cvtColor(hsv.astype(np.uint8), cv2.COLOR_HSV2BGR)
# 4. Sharpness (銳利度)
if sharp_val > 0:
kernel = np.array([[-1, -1, -1], [-1, 9, -1], [-1, -1, -1]])
sharpened = cv2.filter2D(img_bgr, -1, kernel)
img_bgr = cv2.addWeighted(img_bgr, 1.0 - sharp_val, sharpened, sharp_val, 0)
# --- 裁切出特定區域 (ROI) ---
# 只把這個區域送給 AI 辨識，可以提升速度並限定偵測範圍
roi_img = img_bgr[ROI_Y1:ROI_Y2, ROI_X1:ROI_X2]
# --- AI 偵測 (針對 ROI) ---
blob = cv2.dnn.blobFromImage(cv2.resize(roi_img, (300, 300)), 1.0, (300, 300), (104.0, 177.0, 123.0))
net.setInput(blob)
detections = net.forward()
face_count = 0
# 畫出偵測區域的框線 (藍色)
cv2.rectangle(img_bgr, (ROI_X1, ROI_Y1), (ROI_X2, ROI_Y2), (255, 0, 0), 2)
cv2.putText(img_bgr, "Detection Area", (ROI_X1, ROI_Y1 - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (255, 0, 0), 2)
for i in range(0, detections.shape[2]):
confidence = detections[0, 0, i, 2]
if confidence > 0.6:
face_count += 1
# 計算在 ROI 內的座標
box = detections[0, 0, i, 3:7] * np.array([ROI_W, ROI_H, ROI_W, ROI_H])
(startX, startY, endX, endY) = box.astype("int")
# 將 ROI 的座標轉換回「原始全畫面」的座標，才能畫在正確位置
startX += ROI_X1
startY += ROI_Y1
endX += ROI_X1
endY += ROI_Y1
text = f"{confidence * 100:.2f}%"
y = startY - 10 if startY - 10 > 10 else startY + 10
cv2.rectangle(img_bgr, (startX, startY), (endX, endY), (0, 255, 0), 2)
cv2.putText(img_bgr, text, (startX, y), cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
current_time = time.time()
# --- 判斷與顯示狀態 (狀態機防閃爍機制) ---
if face_count > 0:
status_text = f"Status: Face Detected (Count: {face_count})"
status_color = (0, 255, 0) # 綠色
# 狀態機：如果原本沒有人臉，現在偵測到了 -> 觸發語音
if not is_face_present:
is_face_present = True
print(">>> 🟢 狀態切換：進入 ROI，偵測到人臉！")
speech_queue.put("偵測到人臉")
# 更新最後看到人臉的時間
last_face_time = current_time
else:
status_text = "Status: No Face"
status_color = (0, 0, 255) # 紅色
# 狀態機：如果原本有人臉，且離開畫面已經超過 FACE_TIMEOUT 秒 -> 重置狀態
if is_face_present and (current_time - last_face_time > FACE_TIMEOUT):
is_face_present = False
print(f">>> 🔴 狀態切換：人臉已離開 ROI 超過 {FACE_TIMEOUT} 秒，狀態重置！")
# 顯示狀態與人臉個數
cv2.putText(img_bgr, status_text, (20, 90), cv2.FONT_HERSHEY_SIMPLEX, 1.0, status_color, 3)
# 顯示 FPS
cv2.putText(img_bgr, f"FPS: {int(fps)}", (20, 50), cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0, 255, 255), 3)
# --- 發送串流 ---
data = img_bgr.tobytes()
gst_buffer = Gst.Buffer.new_allocate(None, len(data), None)
gst_buffer.fill(0, data)
gst_buffer.pts = buf.pts
gst_buffer.dts = buf.dts
gst_buffer.duration = buf.duration
sender_src.emit("push-buffer", gst_buffer)
# 本地顯示 (整合在同一個 UI)
small_frame = cv2.resize(img_bgr, None, fx=0.5, fy=0.5)
cv2.imshow(UI_WINDOW_NAME, small_frame)
cv2.waitKey(1)
buf.unmap(mapinfo)
return Gst.FlowReturn.OK
# --- 6. 建立攝影機接收管線 ---
cmd = """
mfvideosrc device-index=0 ! decodebin ! videoconvert ! videoscale ! video/x-raw, width=1280, height=720, format=RGB ! appsink name=mysink emit-signals=True drop=True
"""
pipeline = Gst.parse_launch(cmd)
appsink = pipeline.get_by_name("mysink")
appsink.connect("new-sample", on_new_sample)
print("🚀 系統啟動中... (按 Ctrl+C 停止)")
pipeline.set_state(Gst.State.PLAYING)
loop = GLib.MainLoop()
try:
loop.run()
except KeyboardInterrupt:
print("停止中...")
pipeline.set_state(Gst.State.NULL)
sender_pipeline.set_state(Gst.State.NULL)
speech_queue.put(None) # 關閉語音執行緒
cv2.destroyAllWindows()

根據在上一次的
face detect 後，與公司AI部門的同事閒聊後，學到可以設定ROI來設定想要辨識的區域
於是自己回家練習一下，與GEMINI對話後，請 GEMINI base在前一版的基礎下，做到以下功能:

之前曾經有一個需求，就是能在所有監視器的影片存檔中，能有工具能找出出現特定物件的時間，與存檔檔名這次利用空檔，請gemini寫了一個 python程式，程式功能如下:--
.UI能讓使用者選擇影片的目錄.可以指令想要搜尋的目標，目前有狗、貓、腳踏車、人.如果有顯卡如 RTX 4060就會利用GPU來加速處理.支援多執行序列.搜尋過程會秀出發現的每個時間點.最後如果有搜尋到，會產生一個excel報告檔案，提供給使用者參考
--

趁著有空檔，使用 gemini,請 gemini教我使用 gstreamer 與 gst-launch-1.0於是就問gemini一個主題，請他寫一個python程式，可以利用 winodws 11上面的 usb camera來做一下人臉辨識的程式也就是 在 Windows 11 上成功跑通 GStreamer + Python + Webcam + AI 推論迴圈目前的功能為.用現代的 深度學習 (Deep Learning, DNN) 模型
deploy.prototxt (模型結構描述)
res10_300x300_ssd_iter_140000.caffemodel (訓練好的權重).辨識人臉，並加上框框，並秀出辨識信心度

在個人研究上，有時候會有很多pdf文件需要搜尋，且不知道是在哪個pdf文件時，就很需要一個pdf全文index搜尋的程式有了需求，就試用一下gemini，看看gemini能否寫一個python的程式來達到我的需求於是下完一些prompt與修正後，就有了以下成品環境、安裝與使用方式如下:--environment-
.windows 11
.python --version , Python 3.12.4
//install whoosh and pypdf
pip install whoosh pypdf
//run application
python pdf_search_app.py

這次要利用gemini幫我設計一個賓果遊戲，使用方式如下
程式執行環境
----
.windows 11
.python --version , Python 3.12.4
.pip install flask
.python bingo_party.py
遊戲方法
-----
.掃描QRcode後，輸入姓名、性別後加入。
.等待特參賽者mark-yang主控端來啟動遊戲。
.亂數產生1-25數字棋盤。
.輪流由每個參賽者選擇數字。
.最先連線三條線的獲勝。
.選出前三名優勝者。
-----

今天問AI要如何學習gstreamer與 gst-launch-1.0相關知識。Gemini -> [引導式學習]，教我很多東西，以下做紀錄。
PROMPT:    " 我對於 gstreamer不太熟悉，可以先淺顯的教我 gstreamer與 gst-launch-1.0相關知識嗎? "
GEMINI:    歷史對話
    https://gemini.google.com/share/e3bf1d5f2a36

我利用 gemini 的 canvas用prompt寫了一個"羽球分組對戰產生器"這個羽球分組對戰產生器的功用如下:===============羽球分組對戰產生器
https://gemini.google.com/share/c930458ecaa4
1.輸入所有人的姓名
2.選隊長，區分兩隊
3.其餘人員隨機分隊
4.產生出對戰表

以下為運動員公仔圖像生成 語法與示意圖:上傳的照片檔案為主角
A full-figure action figure of a male display in its original blister pack packaging. 3D toy style. On the top packaging cards, the name MArk Yang is written prominently, with the role badmintion player written below it. the figure represenrs Mark Yang is wearing yonex badmintion sports set , sports shoes. inside the blister pack , next to the figure , are the following accessories: 2 yonex badmintion rackets, yones badmintion balls, specialized badmintion box ,specialized badmintion backpack, racket grip, sport water bottle.Packaging card has a sports theme,with a badminton image , bright background color with a touch of bright pink and sky.photorealistic rendening , studio lighting , clear focus on th packageing and figure. --ar2:3