perf(video): 优化视频处理性能监控和音频播放

- 添加视频处理性能计时和统计功能
- 实现帧处理时间监控和慢帧警告
- 添加音频文件静音修剪功能
- 优化Windows平台音频播放实现
- 调整默认日志输出频率减少冗余信息
- 修复MediaPipe GPU委托在Windows上的兼容性问题

.gitignore

@@ -2,3 +2,5 @@
 .idea/
 __pycache__/
 logs/
+
+resources/

app/audio/generate.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 import platform
 import shutil
 import subprocess
+import wave
 from pathlib import Path
 
 from loguru import logger
@@ -15,54 +16,65 @@ def generate_rep_audio_files(
     rate: int,
     output_dir: Path,
     overwrite: bool = False,
+    trim_leading_silence: bool = True,
+    trim_silence_threshold: int = 500,
+    trim_silence_padding_ms: int = 20,
 ) -> None:
     """
     确保 0~max_count 的运动次数语音 wav 文件存在。
 
     默认生成到：
 
-        app/audio/reps/0.wav
-        app/audio/reps/1.wav
+        resources/audio/reps/0.aiff  # macOS
+        resources/audio/reps/0.wav   # Windows / Linux
         ...
-        app/audio/reps/200.wav
+        resources/audio/reps/200.aiff 或 200.wav
 
     服务启动时调用一次即可。
     """
     output_dir.mkdir(parents=True, exist_ok=True)
+    system = platform.system().lower()
+    suffix = ".aiff" if system == "darwin" else ".wav"
 
     missing_counts = [
         count
         for count in range(0, max_count + 1)
-        if overwrite or not _audio_path(output_dir, count).exists()
+        if overwrite or not _audio_path(output_dir, count, suffix=suffix).exists()
     ]
 
     if not missing_counts:
         logger.info("Rep audio files already prepared: {}", output_dir)
-        return
-
-    system = platform.system().lower()
-
-    logger.info(
-        "Preparing rep audio files, system={}, count={}, output_dir={}",
-        system,
-        len(missing_counts),
-        output_dir,
-    )
-
-    if system == "darwin":
-        _generate_with_macos_say(
-            counts=missing_counts,
-            output_dir=output_dir,
-            rate=rate,
-        )
     else:
-        _generate_with_pyttsx3(
-            counts=missing_counts,
-            output_dir=output_dir,
-            rate=rate,
+        logger.info(
+            "Preparing rep audio files, system={}, count={}, output_dir={}",
+            system,
+            len(missing_counts),
+            output_dir,
         )
 
-    logger.info("Rep audio files prepared: {}", output_dir)
+        if system == "darwin":
+            _generate_with_macos_say(
+                counts=missing_counts,
+                output_dir=output_dir,
+                rate=rate,
+            )
+        else:
+            _generate_with_pyttsx3(
+                counts=missing_counts,
+                output_dir=output_dir,
+                rate=rate,
+            )
+
+        logger.info("Rep audio files prepared: {}", output_dir)
+
+    if trim_leading_silence and suffix == ".wav":
+        _trim_leading_silence_files(
+            counts=list(range(0, max_count + 1)),
+            output_dir=output_dir,
+            suffix=suffix,
+            threshold=trim_silence_threshold,
+            padding_ms=trim_silence_padding_ms,
+        )
 
 
 def _generate_with_macos_say(
@@ -79,22 +91,27 @@ def _generate_with_macos_say(
         raise RuntimeError("macOS say command not found")
 
     for count in counts:
-        audio_file = _audio_path(output_dir, count)
+        audio_file = _audio_path(output_dir, count, suffix=".aiff")
 
-        subprocess.run(
-            [
-                "say",
-                "-r",
-                str(rate),
-                "--file-format=WAVE",
-                "-o",
-                str(audio_file),
-                str(count),
-            ],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-            check=True,
-        )
+        try:
+            subprocess.run(
+                [
+                    "say",
+                    "-r",
+                    str(rate),
+                    "--file-format=AIFF",
+                    "-o",
+                    str(audio_file),
+                    str(count),
+                ],
+                stdout=subprocess.DEVNULL,
+                stderr=subprocess.PIPE,
+                text=True,
+                check=True,
+            )
+        except subprocess.CalledProcessError as exc:
+            message = exc.stderr.strip() or f"exit status {exc.returncode}"
+            raise RuntimeError(f"Failed to generate {audio_file}: {message}") from exc
 
 
 def _generate_with_pyttsx3(
@@ -114,11 +131,97 @@ def _generate_with_pyttsx3(
     engine.setProperty("volume", 1.0)
 
     for count in counts:
-        audio_file = _audio_path(output_dir, count)
+        audio_file = _audio_path(output_dir, count, suffix=".wav")
         engine.save_to_file(str(count), str(audio_file))
 
     engine.runAndWait()
 
 
-def _audio_path(output_dir: Path, count: int) -> Path:
-    return output_dir / f"{count}.wav"
+def _audio_path(output_dir: Path, count: int, *, suffix: str) -> Path:
+    return output_dir / f"{count}{suffix}"
+
+
+def _trim_leading_silence_files(
+    *,
+    counts: list[int],
+    output_dir: Path,
+    suffix: str,
+    threshold: int,
+    padding_ms: int,
+) -> None:
+    trimmed = 0
+    total_removed_ms = 0.0
+
+    for count in counts:
+        audio_file = _audio_path(output_dir, count, suffix=suffix)
+        if not audio_file.exists():
+            continue
+        removed_ms = _trim_leading_silence(audio_file, threshold=threshold, padding_ms=padding_ms)
+        if removed_ms > 0:
+            trimmed += 1
+            total_removed_ms += removed_ms
+
+    logger.info(
+        "Rep audio leading silence trim complete: files_trimmed={}, total_removed_ms={:.1f}, threshold={}, padding_ms={}",
+        trimmed,
+        total_removed_ms,
+        threshold,
+        padding_ms,
+    )
+
+
+def _trim_leading_silence(audio_file: Path, *, threshold: int, padding_ms: int) -> float:
+    with wave.open(str(audio_file), "rb") as reader:
+        params = reader.getparams()
+        frames = reader.readframes(params.nframes)
+
+    frame_size = params.sampwidth * params.nchannels
+    if params.nframes <= 0 or frame_size <= 0:
+        return 0.0
+
+    chunk_frames = max(1, params.framerate // 100)
+    leading_frames = 0
+    offset = 0
+    chunk_size = chunk_frames * frame_size
+
+    while offset < len(frames):
+        chunk = frames[offset : offset + chunk_size]
+        if _pcm_rms(chunk, params.sampwidth) > threshold:
+            break
+        chunk_frame_count = len(chunk) // frame_size
+        leading_frames += chunk_frame_count
+        offset += chunk_size
+
+    padding_frames = int(params.framerate * max(0, padding_ms) / 1000)
+    remove_frames = max(0, leading_frames - padding_frames)
+    if remove_frames <= 0:
+        return 0.0
+
+    start = min(len(frames), remove_frames * frame_size)
+    trimmed_frames = frames[start:]
+    if not trimmed_frames:
+        return 0.0
+
+    with wave.open(str(audio_file), "wb") as writer:
+        writer.setparams(params)
+        writer.writeframes(trimmed_frames)
+
+    return remove_frames / params.framerate * 1000
+
+
+def _pcm_rms(chunk: bytes, sample_width: int) -> float:
+    if not chunk:
+        return 0.0
+
+    if sample_width == 2:
+        sample_count = len(chunk) // 2
+        if sample_count == 0:
+            return 0.0
+        total = 0
+        for i in range(0, sample_count * 2, 2):
+            sample = int.from_bytes(chunk[i : i + 2], "little", signed=True)
+            total += sample * sample
+        return (total / sample_count) ** 0.5
+
+    peak = max(abs(byte - 128) for byte in chunk)
+    return float(peak)

app/audio/rep_announcer.py

@@ -1,43 +1,38 @@
 from __future__ import annotations
 
-import os
 import queue
 import shutil
 import subprocess
 import sys
 import threading
+import time
 from pathlib import Path
-from typing import Any
 
 from loguru import logger
 
 
 class RepAnnouncer:
-    """运动次数语音播报器：预生成 0~200 音频文件，运行时直接播放"""
+    """运动次数语音播报器：读取预生成的音频文件直接播放"""
 
     def __init__(
         self,
         *,
         enabled: bool = True,
-        rate: int = 185,
-        volume: float = 1.0,
         max_count: int = 200,
-        cache_dir: str | Path = "runtime/tts_cache/reps",
+        audio_dir: str | Path = "resources/audio/reps",
     ) -> None:
         self.enabled = enabled
-        self.rate = rate
-        self.volume = volume
         self.max_count = max_count
-        self.cache_dir = Path(cache_dir)
+        self.audio_dir = Path(audio_dir)
 
-        self._queue: queue.Queue[int | None] = queue.Queue()
+        self._queue: queue.Queue[tuple[int, float] | None] = queue.Queue()
         self._thread: threading.Thread | None = None
-        self._engine: Any | None = None
         self._current_process: subprocess.Popen | None = None
         self._closed = False
+        self._play_lock = threading.Lock()
 
-        self._use_macos_say = sys.platform == "darwin"
-        self._use_windows_winsound = sys.platform.startswith("win")
+        self._platform = sys.platform
+        self._direct_playback = self._platform.startswith("win")
 
         if self.enabled:
             self._start()
@@ -46,13 +41,29 @@ class RepAnnouncer:
         """将次数放入队列，后台线程播放对应音频"""
         if not self.enabled or self._closed:
             return
-
         if count <= 0 or count > self.max_count:
             return
 
-        # 保留“只播最新一次”的策略，避免语音堆积
+        requested_at = time.perf_counter()
+        if self._direct_playback:
+            audio_file = self._audio_path(count)
+            if not audio_file.exists():
+                logger.warning("Rep audio file missing: {}", audio_file)
+                return
+            try:
+                self._play(audio_file)
+                logger.info(
+                    "Rep audio submitted immediately: count={}, submit_ms={:.1f}",
+                    count,
+                    (time.perf_counter() - requested_at) * 1000,
+                )
+            except Exception as exc:
+                logger.warning("Failed to play rep count {}: {}", count, exc)
+            return
+
         self._clear_pending_counts()
-        self._queue.put(count)
+        self._queue.put((count, requested_at))
+        logger.info("Rep audio queued: count={}", count)
 
     def close(self) -> None:
         """停止播报线程并释放资源"""
@@ -66,145 +77,78 @@ class RepAnnouncer:
             self._thread.join(timeout=1.0)
 
         self._stop_current_playback()
-
         logger.info("Rep announcer closed")
 
     def _start(self) -> None:
-        """初始化并预生成语音缓存"""
-        self.cache_dir.mkdir(parents=True, exist_ok=True)
+        """启动后台播报线程"""
+        self.audio_dir.mkdir(parents=True, exist_ok=True)
 
-        try:
-            self._prepare_audio_cache()
-        except Exception as exc:
-            self.enabled = False
-            logger.warning("Rep announcer disabled, failed to prepare audio cache: {}", exc)
+        if self._direct_playback:
+            import winsound
+
+            logger.info("Rep announcer initialized in direct Windows mode, audio_dir={}", self.audio_dir)
             return
 
-        self._thread = threading.Thread(
-            target=self._run,
-            name="RepAnnouncer",
-            daemon=True,
-        )
+        self._thread = threading.Thread(target=self._run, name="RepAnnouncer", daemon=True)
         self._thread.start()
 
-        logger.info(
-            "Rep announcer initialized with audio cache, platform={}, max_count={}, cache_dir={}",
-            sys.platform,
-            self.max_count,
-            self.cache_dir,
-        )
-
-    def _prepare_audio_cache(self) -> None:
-        """生成 0~max_count 的语音文件，只生成缺失文件"""
-        if self._use_macos_say:
-            self._prepare_macos_say_cache()
-        else:
-            self._prepare_pyttsx3_cache()
-
-    def _prepare_macos_say_cache(self) -> None:
-        """macOS: 使用 say 预生成 aiff 文件"""
-        if shutil.which("say") is None:
-            raise RuntimeError("macOS say command not found")
-
-        for count in range(0, self.max_count + 1):
-            audio_file = self._audio_path(count)
-
-            if audio_file.exists():
-                continue
-
-            subprocess.run(
-                [
-                    "say",
-                    "-r",
-                    str(self.rate),
-                    "-o",
-                    str(audio_file),
-                    str(count),
-                ],
-                stdout=subprocess.DEVNULL,
-                stderr=subprocess.DEVNULL,
-                check=True,
-            )
-
-        logger.info("macOS say audio cache prepared")
-
-    def _prepare_pyttsx3_cache(self) -> None:
-        """非 macOS: 使用 pyttsx3 预生成 wav 文件"""
-        try:
-            import pyttsx3
-        except Exception as exc:
-            raise RuntimeError(f"pyttsx3 unavailable: {exc}") from exc
-
-        self._engine = pyttsx3.init()
-        self._engine.setProperty("rate", self.rate)
-        self._engine.setProperty("volume", self.volume)
-
-        need_generate = False
-
-        for count in range(0, self.max_count + 1):
-            if not self._audio_path(count).exists():
-                need_generate = True
-                self._engine.save_to_file(str(count), str(self._audio_path(count)))
-
-        if need_generate:
-            self._engine.runAndWait()
-
-        logger.info("pyttsx3 audio cache prepared")
+        logger.info("Rep announcer initialized in queued mode, audio_dir={}", self.audio_dir)
 
     def _run(self) -> None:
-        """后台线程：只负责播放已经生成好的音频文件"""
+        """后台线程：从队列取次数，播放对应音频文件"""
         while True:
-            count = self._queue.get()
-
-            if count is None:
+            item = self._queue.get()
+            if item is None:
                 return
+            count, requested_at = item
+
+            audio_file = self._audio_path(count)
+            if not audio_file.exists():
+                logger.warning("Rep audio file missing: {}", audio_file)
+                continue
 
             try:
-                audio_file = self._audio_path(count)
-
-                if not audio_file.exists():
-                    logger.warning("Rep audio file missing: {}", audio_file)
-                    continue
-
-                self._play_audio(audio_file)
-
+                self._play(audio_file)
+                logger.info(
+                    "Rep audio submitted from queue: count={}, queue_ms={:.1f}",
+                    count,
+                    (time.perf_counter() - requested_at) * 1000,
+                )
             except Exception as exc:
                 logger.warning("Failed to play rep count {}: {}", count, exc)
 
-    def _play_audio(self, audio_file: Path) -> None:
-        """根据平台播放音频"""
-        self._stop_current_playback()
+    def _play(self, audio_file: Path) -> None:
+        """播放音频文件（平台自适应）"""
+        with self._play_lock:
+            self._stop_current_playback()
 
-        if self._use_macos_say:
-            self._current_process = subprocess.Popen(
-                ["afplay", str(audio_file)],
-                stdout=subprocess.DEVNULL,
-                stderr=subprocess.DEVNULL,
-            )
-            return
+            if self._platform == "darwin":
+                self._current_process = subprocess.Popen(
+                    ["afplay", str(audio_file)],
+                    stdout=subprocess.DEVNULL,
+                    stderr=subprocess.DEVNULL,
+                )
+            elif self._platform.startswith("win"):
+                import winsound
 
-        if self._use_windows_winsound:
-            import winsound
-
-            # SND_ASYNC 表示异步播放；PURGE 会被 _stop_current_playback 调用中断
-            winsound.PlaySound(str(audio_file), winsound.SND_FILENAME | winsound.SND_ASYNC)
-            return
-
-        # Linux：优先 paplay，其次 aplay
-        player = shutil.which("paplay") or shutil.which("aplay")
-        if player is None:
-            logger.warning("No audio player found, expected paplay or aplay")
-            return
-
-        self._current_process = subprocess.Popen(
-            [player, str(audio_file)],
-            stdout=subprocess.DEVNULL,
-            stderr=subprocess.DEVNULL,
-        )
+                winsound.PlaySound(
+                    str(audio_file),
+                    winsound.SND_FILENAME | winsound.SND_ASYNC | winsound.SND_NODEFAULT,
+                )
+            else:
+                player = shutil.which("paplay") or shutil.which("aplay")
+                if player is None:
+                    logger.warning("No audio player found")
+                    return
+                self._current_process = subprocess.Popen(
+                    [player, str(audio_file)],
+                    stdout=subprocess.DEVNULL,
+                    stderr=subprocess.DEVNULL,
+                )
 
     def _stop_current_playback(self) -> None:
         """中断当前正在播放的声音"""
-        if self._use_windows_winsound:
+        if self._platform.startswith("win"):
             try:
                 import winsound
 
@@ -215,13 +159,12 @@ class RepAnnouncer:
 
         if self._current_process is not None and self._current_process.poll() is None:
             self._current_process.terminate()
-
         self._current_process = None
 
     def _audio_path(self, count: int) -> Path:
         """获取某个次数对应的音频文件路径"""
-        suffix = ".aiff" if self._use_macos_say else ".wav"
-        return self.cache_dir / f"{count}{suffix}"
+        suffix = ".aiff" if self._platform == "darwin" else ".wav"
+        return self.audio_dir / f"{count}{suffix}"
 
     def _clear_pending_counts(self) -> None:
         """清空队列中等待播放的次数，避免语音堆积"""
@@ -229,7 +172,6 @@ class RepAnnouncer:
             try:
                 item = self._queue.get_nowait()
                 if item is None:
-                    # close 信号不要吞掉
                     self._queue.put(None)
                     return
             except queue.Empty:

app/core/lifecycle.py

@@ -2,10 +2,21 @@ from __future__ import annotations
 
 from app.diagnostics.crash_handler import enable_crash_handler
 from configs.load import config
-
+from app.audio.generate import generate_rep_audio_files
 
 def startup() -> None:
     """应用启动初始化：开启崩溃日志和日志系统"""
     enable_crash_handler(config.logging.dir_path)
     from app.core.logging import setup_logging
     setup_logging()
+
+    # 生成运动次数语音文件
+    generate_rep_audio_files(
+        max_count=config.audio.rep_max_count,
+        rate=config.audio.rep_announcer_rate,
+        output_dir=config.audio.resolved_audio_dir,
+        overwrite=False,
+        trim_leading_silence=config.audio.trim_leading_silence,
+        trim_silence_threshold=config.audio.trim_silence_threshold,
+        trim_silence_padding_ms=config.audio.trim_silence_padding_ms,
+    )

app/exercises/dead_bug/detector.py

@@ -52,6 +52,7 @@ class DeadBugDetector:
         self._result_event = threading.Event()
         self._inflight = False
         self._inflight_started_at = 0.0
+        self.last_timing: dict[str, float | bool] = {}
 
         def on_result(pose_result, _image, _timestamp_ms):
             with self._result_lock:
@@ -79,7 +80,9 @@ class DeadBugDetector:
 
     def process_frame(self, bgr_frame: np.ndarray, timestamp_ms: int) -> tuple[np.ndarray, DeadBugResult]:
         """处理单帧：姿态检测、指标计算、状态机更新、可视化叠加"""
+        total_started = time.perf_counter()
         timestamp_ms = self._normalize_timestamp(timestamp_ms)
+        normalize_done = time.perf_counter()
 
         with self._result_lock:
             if self._inflight and time.monotonic() - self._inflight_started_at > 0.5:
@@ -90,9 +93,11 @@ class DeadBugDetector:
             if should_submit:
                 self._inflight = True
                 self._inflight_started_at = time.monotonic()
+        lock_done = time.perf_counter()
 
         if should_submit:
             rgba_frame = bgr_to_rgba(bgr_frame)
+            convert_done = time.perf_counter()
             mp_image = mp.Image(image_format=mp.ImageFormat.SRGBA, data=rgba_frame)
             self._result_event.clear()
             try:
@@ -102,14 +107,23 @@ class DeadBugDetector:
                     self._inflight = False
                     self._inflight_started_at = 0.0
                 raise
+            submit_done = time.perf_counter()
             self._result_event.wait(timeout=0.08)
+            wait_done = time.perf_counter()
+        else:
+            convert_done = lock_done
+            submit_done = lock_done
+            wait_done = lock_done
 
         with self._result_lock:
             pose_result = self._latest_result
+        result_read_done = time.perf_counter()
 
         annotated = bgr_frame.copy()
+        copy_done = time.perf_counter()
 
         if pose_result is None or not pose_result.pose_landmarks:
+            self._state.mark_no_pose()
             result = DeadBugResult(
                 rep_count=self._state.rep_count,
                 phase=DeadBugPhase.NO_POSE,
@@ -119,12 +133,25 @@ class DeadBugDetector:
                 metrics=None,
             )
             draw_status_overlay(annotated, result)
+            self._record_timing(
+                total_started,
+                normalize_done,
+                lock_done,
+                convert_done,
+                submit_done,
+                wait_done,
+                result_read_done,
+                copy_done,
+                time.perf_counter(),
+                should_submit,
+            )
             return annotated, result
 
         landmarks = [Point(lm.x, lm.y, lm.z, getattr(lm, "visibility", 1.0)) for lm in pose_result.pose_landmarks[0]]
         draw_landmarks(annotated, landmarks, REQUIRED_LANDMARKS, visibility_threshold=self.visibility_threshold)
 
         if not has_required_visibility(landmarks, REQUIRED_LANDMARKS, self.visibility_threshold):
+            self._state.mark_no_pose()
             result = DeadBugResult(
                 rep_count=self._state.rep_count,
                 phase=DeadBugPhase.NO_POSE,
@@ -134,6 +161,18 @@ class DeadBugDetector:
                 metrics=None,
             )
             draw_status_overlay(annotated, result)
+            self._record_timing(
+                total_started,
+                normalize_done,
+                lock_done,
+                convert_done,
+                submit_done,
+                wait_done,
+                result_read_done,
+                copy_done,
+                time.perf_counter(),
+                should_submit,
+            )
             return annotated, result
 
         raw = calculate_metrics(
@@ -167,8 +206,46 @@ class DeadBugDetector:
 
         result = self._state.update(metrics)
         draw_status_overlay(annotated, result)
+        self._record_timing(
+            total_started,
+            normalize_done,
+            lock_done,
+            convert_done,
+            submit_done,
+            wait_done,
+            result_read_done,
+            copy_done,
+            time.perf_counter(),
+            should_submit,
+        )
         return annotated, result
 
+    def _record_timing(
+        self,
+        total_started: float,
+        normalize_done: float,
+        lock_done: float,
+        convert_done: float,
+        submit_done: float,
+        wait_done: float,
+        result_read_done: float,
+        copy_done: float,
+        finished: float,
+        submitted: bool,
+    ) -> None:
+        self.last_timing = {
+            "total_ms": (finished - total_started) * 1000,
+            "timestamp_ms": (normalize_done - total_started) * 1000,
+            "lock_ms": (lock_done - normalize_done) * 1000,
+            "convert_ms": (convert_done - lock_done) * 1000,
+            "submit_ms": (submit_done - convert_done) * 1000,
+            "wait_ms": (wait_done - submit_done) * 1000,
+            "result_read_ms": (result_read_done - wait_done) * 1000,
+            "copy_ms": (copy_done - result_read_done) * 1000,
+            "postprocess_draw_ms": (finished - copy_done) * 1000,
+            "submitted": submitted,
+        }
+
     def _normalize_timestamp(self, timestamp_ms: int) -> int:
         """确保时间戳严格递增（MediaPipe要求）"""
         if timestamp_ms <= self._last_timestamp_ms:

app/exercises/dead_bug/rules.py

@@ -9,7 +9,7 @@ def has_required_visibility(landmarks: list[Point], required_indices: tuple[int,
 
 
 def detect_diagonal_extension(metrics: DeadBugMetrics) -> str | None:
-    """检测是否存在对角伸展（左臂+右腿 或 右臂+左腿）"""
+    """检测对角伸展（腿部只允许单侧伸展，手臂允许准备位上举带来的识别重叠）"""
     if metrics.left_leg_extended and metrics.right_leg_extended:
         return None
 
@@ -21,7 +21,7 @@ def detect_diagonal_extension(metrics: DeadBugMetrics) -> str | None:
 
 
 def is_ready_position(metrics: DeadBugMetrics) -> bool:
-    """判断是否处于准备姿态（膝盖弯曲且四肢未伸展）"""
+    """判断是否处于准备姿态（双膝弯曲且双腿未伸展；dead bug 准备位允许手臂上举）"""
     knees_bent = metrics.left_knee_angle <= 140 and metrics.right_knee_angle <= 140
     legs_not_extended = not metrics.left_leg_extended and not metrics.right_leg_extended
-    return knees_bent and legs_not_extended and detect_diagonal_extension(metrics) is None
+    return knees_bent and legs_not_extended

app/exercises/dead_bug/state_machine.py

@@ -3,6 +3,13 @@ from __future__ import annotations
 from app.exercises.dead_bug.rules import detect_diagonal_extension, is_ready_position
 from app.exercises.dead_bug.types import DeadBugMetrics, DeadBugPhase, DeadBugResult
 
+_SMOOTHING_ALPHA = 0.45
+_TREND_DELTA_DEGREES = 1.5
+_SIDE_ANGLE_MARGIN = 10.0
+_EXTENSION_START_ANGLE = 125.0
+_EXTENSION_PEAK_ANGLE = 150.0
+_READY_KNEE_ANGLE = 140.0
+
 class DeadBugStateMachine:
     """死虫式动作状态机：管理READY/EXTENDING/NEED_RESET/NO_POSE状态转换"""
 
@@ -17,11 +24,26 @@ class DeadBugStateMachine:
         self._candidate_side: str | None = None
         self._candidate_frames = 0
         self._reset_frames = 0
+        self._smooth_left_knee_angle: float | None = None
+        self._smooth_right_knee_angle: float | None = None
+        self._left_knee_delta = 0.0
+        self._right_knee_delta = 0.0
+
+    def mark_no_pose(self) -> None:
+        """姿态丢失时清掉候选帧；已确认的半程动作保留，等待重新可见后完成回收。"""
+        if self.phase == DeadBugPhase.READY:
+            self.phase = DeadBugPhase.NO_POSE
+            self.active_side = None
+        self._candidate_side = None
+        self._candidate_frames = 0
+        self._reset_frames = 0
 
     def update(self, metrics: DeadBugMetrics) -> DeadBugResult:
         """根据传入指标更新状态机并返回本次结果"""
-        side = detect_diagonal_extension(metrics)
-        ready = is_ready_position(metrics)
+        self._update_knee_trends(metrics)
+
+        side = self._detect_motion_side(metrics)
+        ready = self._is_stable_ready(metrics)
 
         if side is None:
             self._candidate_side = None
@@ -33,13 +55,16 @@ class DeadBugStateMachine:
             self._candidate_frames = 1
 
         if self.phase in (DeadBugPhase.READY, DeadBugPhase.NO_POSE):
+            if ready:
+                self.phase = DeadBugPhase.READY
             if self._candidate_frames >= self.extension_confirm_frames and side is not None:
                 self.phase = DeadBugPhase.EXTENDING
                 self.active_side = side
                 self._reset_frames = 0
         elif self.phase == DeadBugPhase.EXTENDING:
-            if side == self.active_side:
+            if ready or self._active_knee_retracting():
                 self.phase = DeadBugPhase.NEED_RESET
+                self._reset_frames = 1 if ready else 0
         elif self.phase == DeadBugPhase.NEED_RESET:
             if ready:
                 self._reset_frames += 1
@@ -54,21 +79,106 @@ class DeadBugStateMachine:
                 self._reset_frames = 0
 
         feedback = list(metrics.feedback)
-        if side is None and not ready:
+        display_side = detect_diagonal_extension(metrics)
+        if display_side is None and not ready:
             feedback.append("Extend opposite arm and leg only")
         if ready:
             feedback.append("Ready position")
-        elif side == "left_arm_right_leg":
+        elif display_side == "left_arm_right_leg":
             feedback.append("Left arm + right leg")
-        elif side == "right_arm_left_leg":
+        elif display_side == "right_arm_left_leg":
             feedback.append("Right arm + left leg")
 
-        is_standard = side is not None and not metrics.feedback
+        is_standard = display_side is not None and not metrics.feedback
         return DeadBugResult(
             rep_count=self.rep_count,
             phase=self.phase,
-            side=side,
+            side=display_side,
             is_standard=is_standard,
             feedback=feedback[:3],
             metrics=metrics,
         )
+
+    def _update_knee_trends(self, metrics: DeadBugMetrics) -> None:
+        """更新平滑膝角和本帧变化量，用连续趋势抵消单帧抖动。"""
+        previous_left = self._smooth_left_knee_angle
+        previous_right = self._smooth_right_knee_angle
+
+        if previous_left is None:
+            self._smooth_left_knee_angle = metrics.left_knee_angle
+            self._left_knee_delta = 0.0
+        else:
+            self._smooth_left_knee_angle = (
+                _SMOOTHING_ALPHA * metrics.left_knee_angle
+                + (1.0 - _SMOOTHING_ALPHA) * previous_left
+            )
+            self._left_knee_delta = self._smooth_left_knee_angle - previous_left
+
+        if previous_right is None:
+            self._smooth_right_knee_angle = metrics.right_knee_angle
+            self._right_knee_delta = 0.0
+        else:
+            self._smooth_right_knee_angle = (
+                _SMOOTHING_ALPHA * metrics.right_knee_angle
+                + (1.0 - _SMOOTHING_ALPHA) * previous_right
+            )
+            self._right_knee_delta = self._smooth_right_knee_angle - previous_right
+
+    def _detect_motion_side(self, metrics: DeadBugMetrics) -> str | None:
+        """基于膝角趋势推断正在伸展的对角侧，手臂只作为辅助校验。"""
+        raw_side = detect_diagonal_extension(metrics)
+        if raw_side is not None and self._side_has_extension_motion(raw_side):
+            return raw_side
+
+        left = self._smooth_left_knee_angle
+        right = self._smooth_right_knee_angle
+        if left is None or right is None:
+            return raw_side
+
+        both_legs_high = left >= _EXTENSION_PEAK_ANGLE and right >= _EXTENSION_PEAK_ANGLE
+        if both_legs_high and abs(left - right) < _SIDE_ANGLE_MARGIN:
+            return None
+
+        if (
+            right >= _EXTENSION_START_ANGLE
+            and right - left >= _SIDE_ANGLE_MARGIN
+            and (self._right_knee_delta >= _TREND_DELTA_DEGREES or right >= _EXTENSION_PEAK_ANGLE)
+        ):
+            return "left_arm_right_leg"
+        if (
+            left >= _EXTENSION_START_ANGLE
+            and left - right >= _SIDE_ANGLE_MARGIN
+            and (self._left_knee_delta >= _TREND_DELTA_DEGREES or left >= _EXTENSION_PEAK_ANGLE)
+        ):
+            return "right_arm_left_leg"
+        return raw_side
+
+    def _side_has_extension_motion(self, side: str) -> bool:
+        """确认对应腿处于伸展区或仍在伸展趋势中。"""
+        if side == "left_arm_right_leg":
+            angle = self._smooth_right_knee_angle
+            delta = self._right_knee_delta
+        else:
+            angle = self._smooth_left_knee_angle
+            delta = self._left_knee_delta
+        if angle is None:
+            return True
+        return angle >= _EXTENSION_PEAK_ANGLE or (
+            angle >= _EXTENSION_START_ANGLE and delta >= _TREND_DELTA_DEGREES
+        )
+
+    def _is_stable_ready(self, metrics: DeadBugMetrics) -> bool:
+        """准备位需要双腿回到屈膝区域；使用平滑膝角避免单帧阈值跳变。"""
+        left = self._smooth_left_knee_angle
+        right = self._smooth_right_knee_angle
+        if left is None or right is None:
+            return is_ready_position(metrics)
+        return left <= _READY_KNEE_ANGLE and right <= _READY_KNEE_ANGLE
+
+    def _active_knee_retracting(self) -> bool:
+        """确认已伸展的那条腿开始回收。"""
+        if self.active_side == "left_arm_right_leg":
+            return self._right_knee_delta <= -_TREND_DELTA_DEGREES
+        if self.active_side == "right_arm_left_leg":
+            return self._left_knee_delta <= -_TREND_DELTA_DEGREES
+        return False

app/vision/pose_landmarker.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import platform
 from typing import Callable
 
 import mediapipe as mp
@@ -27,6 +28,11 @@ class PoseLandmarkerWrapper:
 
         if prefer_gpu:
             try:
+                if platform.system() == "Windows":
+                    logger.warning(
+                        "MediaPipe GPU delegate requested, but MediaPipe Tasks Python does not support GPU delegate on Windows; "
+                        "Intel Iris Xe cannot be used by this backend and CPU fallback is expected"
+                    )
                 self.delegate = BaseOptions.Delegate.GPU
                 self._landmarker = self._create(self.delegate, result_callback)
                 logger.info("MediaPipe PoseLandmarker initialized with GPU delegate")

app/webrtc/video_receiver.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import asyncio
+import time
 
 import cv2
 from aiortc.mediastreams import MediaStreamError
@@ -25,6 +26,39 @@ def _format_pose_debug(pose_result) -> str:
         f"ll={metrics.left_leg_extended}, rl={metrics.right_leg_extended})"
     )
 
+
+def _new_perf_window() -> dict:
+    return {
+        "frames": 0,
+        "processed": 0,
+        "loop_ms": 0.0,
+        "to_ndarray_ms": 0.0,
+        "detect_ms": 0.0,
+        "show_ms": 0.0,
+        "max_loop_ms": 0.0,
+        "max_detect_ms": 0.0,
+        "detector": {},
+    }
+
+
+def _add_detector_timing(perf: dict, timing: dict[str, float | bool]) -> None:
+    detector = perf["detector"]
+    for key, value in timing.items():
+        if key == "submitted":
+            detector[key] = detector.get(key, 0) + (1 if value else 0)
+            continue
+        value = float(value)
+        detector[key] = detector.get(key, 0.0) + value
+        max_key = f"max_{key}"
+        detector[max_key] = max(detector.get(max_key, 0.0), value)
+
+
+def _avg(perf: dict, key: str, denominator: int) -> float:
+    if denominator <= 0:
+        return 0.0
+    return perf.get(key, 0.0) / denominator
+
+
 class VideoReceiver:
     """视频轨道接收与运动检测流水线"""
 
@@ -33,7 +67,21 @@ class VideoReceiver:
 
     async def run(self) -> None:
         """持续接收视频帧并进行姿态检测、渲染和语音播报"""
-        logger.info("Start receiving video frames, process_every_n={}", config.video.process_every_n_frames)
+        log_every_n_frames = max(1, config.video.log_every_n_frames)
+        perf_log_every_n_frames = max(1, config.video.perf_log_every_n_frames)
+        slow_frame_ms = max(0.0, config.video.slow_frame_ms)
+        logger.info(
+            "Start receiving video frames, process_every_n={}, log_every_n={}, perf_log_every_n={}, slow_frame_ms={}",
+            config.video.process_every_n_frames,
+            log_every_n_frames,
+            perf_log_every_n_frames,
+            slow_frame_ms,
+        )
+        logger.info(
+            "OpenCV OpenCL status: have_opencl={}, use_opencl={}",
+            cv2.ocl.haveOpenCL(),
+            cv2.ocl.useOpenCL(),
+        )
 
         frame_count = 0
         processed_count = 0
@@ -46,31 +94,51 @@ class VideoReceiver:
         )
         announcer = RepAnnouncer(
             enabled=config.audio.rep_announcer_enabled,
-            rate=config.audio.rep_announcer_rate,
-            volume=config.audio.rep_announcer_volume,
+            max_count=config.audio.rep_max_count,
+            audio_dir=config.audio.resolved_audio_dir,
         )
         last_announced_rep = 0
         last_pose_result = None
         last_annotated = None
+        perf = _new_perf_window()
 
         try:
             while True:
+                loop_started = time.perf_counter()
                 frame = await self._track.recv()
                 frame_count += 1
+                recv_done = time.perf_counter()
                 raw_img = frame.to_ndarray(format="bgr24")
+                ndarray_done = time.perf_counter()
                 timestamp_ms = int(frame.time * 1000) if frame.time is not None else frame_count * 33
 
+                detect_ms = 0.0
                 if frame_count % config.video.process_every_n_frames == 0 or last_pose_result is None:
+                    detect_started = time.perf_counter()
                     processed_count += 1
                     last_annotated, last_pose_result = detector.process_frame(raw_img, timestamp_ms)
+                    detect_ms = (time.perf_counter() - detect_started) * 1000
+                    perf["processed"] += 1
+                    perf["detect_ms"] += detect_ms
+                    perf["max_detect_ms"] = max(perf["max_detect_ms"], detect_ms)
+                    _add_detector_timing(perf, detector.last_timing)
                     if last_pose_result.rep_count > last_announced_rep:
                         last_announced_rep = last_pose_result.rep_count
+                        announce_started = time.perf_counter()
                         announcer.announce_count(last_announced_rep)
+                        logger.info(
+                            "Rep completed and audio requested: count={}, frame={}, announce_call_ms={:.1f}",
+                            last_announced_rep,
+                            frame_count,
+                            (time.perf_counter() - announce_started) * 1000,
+                        )
 
                 display_img = last_annotated if last_annotated is not None else raw_img
+                show_started = time.perf_counter()
                 show_frame(display_img)
+                show_done = time.perf_counter()
 
-                if frame_count % 100 == 0:
+                if frame_count % log_every_n_frames == 0:
                     logger.info(
                         "Received {} frames, processed={}, raw_shape={}, reps={}, phase={}, feedback={}, {}",
                         frame_count,
@@ -82,6 +150,53 @@ class VideoReceiver:
                         _format_pose_debug(last_pose_result) if last_pose_result is not None else "metrics=None",
                     )
 
+                loop_ms = (show_done - loop_started) * 1000
+                to_ndarray_ms = (ndarray_done - recv_done) * 1000
+                show_ms = (show_done - show_started) * 1000
+                perf["frames"] += 1
+                perf["loop_ms"] += loop_ms
+                perf["to_ndarray_ms"] += to_ndarray_ms
+                perf["show_ms"] += show_ms
+                perf["max_loop_ms"] = max(perf["max_loop_ms"], loop_ms)
+
+                if slow_frame_ms and loop_ms >= slow_frame_ms:
+                    logger.warning(
+                        "Slow video frame: frame={}, loop_ms={:.1f}, detect_ms={:.1f}, to_ndarray_ms={:.1f}, show_ms={:.1f}, shape={}",
+                        frame_count,
+                        loop_ms,
+                        detect_ms,
+                        to_ndarray_ms,
+                        show_ms,
+                        raw_img.shape,
+                    )
+
+                if frame_count % perf_log_every_n_frames == 0:
+                    frames = perf["frames"]
+                    processed = perf["processed"]
+                    detector_perf = perf["detector"]
+                    logger.info(
+                        "Perf window: frames={}, processed={}, avg_loop_ms={:.1f}, max_loop_ms={:.1f}, avg_to_ndarray_ms={:.1f}, "
+                        "avg_detect_ms={:.1f}, max_detect_ms={:.1f}, avg_show_ms={:.1f}, detector_avg_total_ms={:.1f}, "
+                        "detector_max_total_ms={:.1f}, detector_avg_wait_ms={:.1f}, detector_max_wait_ms={:.1f}, "
+                        "detector_avg_convert_ms={:.1f}, detector_avg_postprocess_draw_ms={:.1f}, detector_submitted={}",
+                        frames,
+                        processed,
+                        _avg(perf, "loop_ms", frames),
+                        perf["max_loop_ms"],
+                        _avg(perf, "to_ndarray_ms", frames),
+                        _avg(perf, "detect_ms", processed),
+                        perf["max_detect_ms"],
+                        _avg(perf, "show_ms", frames),
+                        _avg(detector_perf, "total_ms", processed),
+                        detector_perf.get("max_total_ms", 0.0),
+                        _avg(detector_perf, "wait_ms", processed),
+                        detector_perf.get("max_wait_ms", 0.0),
+                        _avg(detector_perf, "convert_ms", processed),
+                        _avg(detector_perf, "postprocess_draw_ms", processed),
+                        detector_perf.get("submitted", 0),
+                    )
+                    perf = _new_perf_window()
+
                 if is_esc_pressed():
                     logger.info("ESC pressed, closing display")
                     break

config.yaml

@@ -7,7 +7,10 @@ server:
   max_ws_size: 10485760   # 10 MB
 
 video:
-  process_every_n_frames: 1
+  process_every_n_frames: 2
+  log_every_n_frames: 30
+  perf_log_every_n_frames: 30
+  slow_frame_ms: 100
 
 model:
   path: "./pose_models/pose_landmarker_full.task"
@@ -24,6 +27,9 @@ audio:
   rep_announcer_volume: 1.0
   rep_max_count: 200              # 预生成语音文件的最大次数
   rep_audio_dir: ""               # 空 = 自动使用 app/audio/reps
+  trim_leading_silence: true
+  trim_silence_threshold: 500
+  trim_silence_padding_ms: 20
 
 logging:
   dir: logs

configs/models.py

@@ -15,7 +15,10 @@ class ServerConfig:
 @dataclass
 class VideoConfig:
     """视频帧处理配置"""
-    process_every_n_frames: int = 1
+    process_every_n_frames: int = 2
+    log_every_n_frames: int = 30
+    perf_log_every_n_frames: int = 30
+    slow_frame_ms: float = 100.0
 
 
 @dataclass
@@ -48,13 +51,16 @@ class AudioConfig:
     rep_announcer_volume: float = 1.0
     rep_max_count: int = 200
     rep_audio_dir: str = ""
+    trim_leading_silence: bool = True
+    trim_silence_threshold: int = 500
+    trim_silence_padding_ms: int = 20
 
     @property
     def resolved_audio_dir(self) -> Path:
         """返回语音文件目录的绝对路径"""
         if self.rep_audio_dir:
             return Path(self.rep_audio_dir)
-        return Path(__file__).resolve().parent.parent / "app" / "audio" / "reps"
+        return Path(__file__).resolve().parent.parent / "resources" / "audio" / "reps"
 
 
 @dataclass

tests/test_dead_bug_rules.py

@@ -49,6 +49,28 @@ class TestDeadBugRules:
         )
         assert detect_diagonal_extension(metrics) == "left_arm_right_leg"
 
+    def test_detect_diagonal_extension_allows_ready_arm_overlap(self):
+        """测试：准备位双臂上举时，单腿对侧伸展仍应识别"""
+        metrics = DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=False, right_leg_extended=True,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=90, right_knee_angle=160,
+            feedback=[],
+        )
+        assert detect_diagonal_extension(metrics) == "left_arm_right_leg"
+
+    def test_detect_diagonal_extension_rejects_both_legs(self):
+        """测试：双腿同时伸展不应识别为可计数对角伸展"""
+        metrics = DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=True, right_leg_extended=True,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=160, right_knee_angle=160,
+            feedback=[],
+        )
+        assert detect_diagonal_extension(metrics) is None
+
     def test_is_ready_position(self):
         """测试：膝盖弯曲且四肢收缩应识别为准备姿态"""
         metrics = DeadBugMetrics(
@@ -60,6 +82,17 @@ class TestDeadBugRules:
         )
         assert is_ready_position(metrics)
 
+    def test_is_ready_allows_arms_extended(self):
+        """测试：dead bug 准备位允许双臂上举"""
+        metrics = DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=False, right_leg_extended=False,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=100, right_knee_angle=100,
+            feedback=[],
+        )
+        assert is_ready_position(metrics)
+
     def test_is_not_ready_legs_extended(self):
         """测试：腿部伸展时不识别为准备姿态"""
         metrics = DeadBugMetrics(

tests/test_dead_bug_state_machine.py

@@ -26,6 +26,36 @@ class TestDeadBugStateMachine:
             feedback=[],
         )
 
+    def _both_legs_extended(self) -> DeadBugMetrics:
+        """构建双腿同时伸展的非标准姿态"""
+        return DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=True, right_leg_extended=True,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=160, right_knee_angle=160,
+            feedback=[],
+        )
+
+    def _arms_extended_ready_legs(self) -> DeadBugMetrics:
+        """构建腿已收回但手臂未收回的姿态"""
+        return DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=False, right_leg_extended=False,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=100, right_knee_angle=100,
+            feedback=[],
+        )
+
+    def _right_knee_angle(self, angle: float) -> DeadBugMetrics:
+        """构建右膝角连续变化但伸展布尔值尚未稳定的姿态"""
+        return DeadBugMetrics(
+            left_arm_extended=True, right_arm_extended=True,
+            left_leg_extended=False, right_leg_extended=False,
+            left_elbow_angle=160, right_elbow_angle=160,
+            left_knee_angle=90, right_knee_angle=angle,
+            feedback=[],
+        )
+
     def test_initial_state(self):
         """测试：状态机初始化后应为READY且计数为0"""
         sm = DeadBugStateMachine()
@@ -46,3 +76,58 @@ class TestDeadBugStateMachine:
         assert sm.phase == DeadBugPhase.READY
         sm.update(self._extended_left())
         assert sm.phase == DeadBugPhase.EXTENDING
+
+    def test_confirm_extension_from_knee_angle_trend(self):
+        """测试：膝角连续上升时，不依赖单帧伸展布尔值也能确认伸展"""
+        sm = DeadBugStateMachine(extension_confirm_frames=2, reset_confirm_frames=2)
+
+        sm.update(self._right_knee_angle(100))
+        sm.update(self._right_knee_angle(130))
+        assert sm.phase == DeadBugPhase.READY
+        sm.update(self._right_knee_angle(145))
+        sm.update(self._right_knee_angle(150))
+
+        assert sm.phase == DeadBugPhase.EXTENDING
+
+    def test_full_rep_counts_once_after_strict_reset(self):
+        """测试：确认伸展后，只有严格回到准备姿态才计一次"""
+        sm = DeadBugStateMachine(extension_confirm_frames=2, reset_confirm_frames=2)
+
+        sm.update(self._extended_left())
+        sm.update(self._extended_left())
+        assert sm.phase == DeadBugPhase.EXTENDING
+
+        sm.update(self._arms_extended_ready_legs())
+        assert sm.rep_count == 0
+        assert sm.phase == DeadBugPhase.NEED_RESET
+
+        sm.update(self._ready_metrics())
+        result = sm.update(self._ready_metrics())
+        assert result.rep_count == 1
+        assert sm.phase == DeadBugPhase.READY
+
+        result = sm.update(self._ready_metrics())
+        assert result.rep_count == 1
+
+    def test_both_legs_do_not_start_rep(self):
+        """测试：双腿同时伸展不进入计数流程"""
+        sm = DeadBugStateMachine(extension_confirm_frames=2, reset_confirm_frames=2)
+        sm.update(self._both_legs_extended())
+        result = sm.update(self._both_legs_extended())
+
+        assert result.rep_count == 0
+        assert sm.phase == DeadBugPhase.READY
+
+    def test_no_pose_preserves_confirmed_rep_until_reset(self):
+        """测试：已确认伸展后短暂丢姿态，回到准备位仍能完成计数"""
+        sm = DeadBugStateMachine(extension_confirm_frames=2, reset_confirm_frames=2)
+        sm.update(self._extended_left())
+        sm.update(self._extended_left())
+        assert sm.phase == DeadBugPhase.EXTENDING
+
+        sm.mark_no_pose()
+        sm.update(self._ready_metrics())
+        result = sm.update(self._ready_metrics())
+
+        assert result.rep_count == 1
+        assert sm.phase == DeadBugPhase.READY