smartpi 0.1.38__py3-none-any.whl

smartpi/rknn_pose_workflow.py

@@ -0,0 +1,592 @@
+import cv2
+import numpy as np
+from PIL import Image
+import json
+import os
+import time
+from rknnlite.api import RKNNLite  # RKNN核心库
+from lib.posenet_utils import get_posenet_output  # 姿态关键点提取逻辑（需支持返回3个值）
+
+
+class PoseWorkflow:
+    def __init__(self, model_path=None):
+        # 基础属性初始化
+        self.model_path = os.path.abspath(model_path) if model_path else None
+        self.rknn_lite = None  # RKNN Lite实例
+        self.classes = []  # 姿态类别标签
+        self.input_shape = []  # 模型输入形状
+        self.output_shape = []  # 模型输出形状
+        self.min_vals = None  # 归一化用最小值
+        self.max_vals = None  # 归一化用最大值
+        
+        # 新增：有效关键点校验配置（可根据需求调整阈值）
+        self.min_valid_keypoints = 1  # 最小有效关键点数量（低于此值则推理终止）
+        
+        # 结果相关属性
+        self.result_image_path = "result.jpg"
+        self.processed_image = None
+        self.last_infer_time = 0.0  # 记录上一帧推理耗时（用于FPS计算）
+
+        # 若传入模型路径，直接加载
+        if model_path:
+            self.load_model()
+
+    def _get_metadata_path(self):
+        """获取元数据文件路径"""
+        if not self.model_path:
+            raise ValueError("模型路径未初始化")
+        
+        # 元数据文件规则：与模型同目录，模型名+"_rknn_metadata.json"
+        base_dir = os.path.dirname(self.model_path)
+        base_name = os.path.basename(self.model_path)
+        metadata_name = os.path.splitext(base_name)[0] + "_rknn_metadata.json"
+        metadata_path = os.path.join(base_dir, metadata_name)
+        
+        # 若自定义名不存在，尝试默认名
+        if not os.path.exists(metadata_path):
+            metadata_path = os.path.join(base_dir, "rknn_metadata.json")
+            print(f"自定义元数据文件不存在，尝试默认路径：{metadata_path}")
+        
+        return metadata_path
+
+    def _load_metadata(self):
+        """加载元数据（classes、input_shape、minMax等）"""
+        metadata_path = self._get_metadata_path()
+        try:
+            with open(metadata_path, "r", encoding="utf-8") as f:
+                metadata = json.load(f)
+            
+            # 读取核心元数据
+            self.classes = metadata.get("classes", [])  # 姿态类别列表
+            self.input_shape = [1, 14739]  # 输入形状，默认14739维
+            self.output_shape = metadata.get("output_shape", [1, len(self.classes)])  # 输出形状
+            min_max = metadata.get("minMax", {})
+            self.min_vals = np.array(min_max.get("min", []), dtype=np.float32)  # 归一化最小值
+            self.max_vals = np.array(min_max.get("max", []), dtype=np.float32)  # 归一化最大值
+            
+        
+        except Exception as e:
+            print(f"元数据加载失败：{e}，使用默认配置")
+            self.classes = []
+            self.input_shape = [1, 14739] 
+            self.min_vals = np.array([])
+            self.max_vals = np.array([])
+
+    def load_model(self):
+        """加载RKNN模型"""
+        try:
+            # 初始化RKNN Lite
+            self.rknn_lite = RKNNLite()
+            
+            # 加载RKNN模型文件
+            ret = self.rknn_lite.load_rknn(self.model_path)
+            if ret != 0:
+                raise RuntimeError(f"RKNN模型加载失败，错误码：{ret}")
+            
+            # 初始化NPU运行时
+            ret = self.rknn_lite.init_runtime()
+            if ret != 0:
+                raise RuntimeError(f"NPU运行时初始化失败，错误码：{ret}")
+            
+            # 加载元数据
+            self._load_metadata()
+            print("RKNN模型加载完成")
+        
+        except Exception as e:
+            print(f"模型加载总失败：{e}")
+            # 释放资源避免泄漏
+            if self.rknn_lite:
+                self.rknn_lite.release()
+
+    def _normalize_pose_points(self, pose_points):
+        """姿态关键点归一化"""
+        pose_points = np.array(pose_points, dtype=np.float32)
+        # 若元数据有minMax且长度匹配，使用minMax归一化
+        if len(self.min_vals) > 0 and len(self.max_vals) > 0 and len(self.min_vals) == len(pose_points):
+            # 避免除以零
+            ranges = self.max_vals - self.min_vals
+            ranges[ranges < 1e-6] = 1e-6
+            normalized = (pose_points - self.min_vals) / ranges
+            return normalized
+        return pose_points
+
+    def _preprocess(self, pose_data):
+        """姿态数据预处理"""
+        try:
+            # 转为numpy数组
+            if not isinstance(pose_data, np.ndarray):
+                pose_data = np.array(pose_data, dtype=np.float32)
+            
+            # 归一化
+            normalized_data = self._normalize_pose_points(pose_data)
+            
+            # 调整为模型输入形状
+            input_size = np.prod(self.input_shape)
+            if normalized_data.size != input_size:
+                print(f"输入数据长度不匹配（实际：{normalized_data.size} | 期望：{input_size}），自动调整维度")
+                normalized_data = np.resize(normalized_data, self.input_shape)
+            else:
+                normalized_data = normalized_data.reshape(self.input_shape)
+            
+            # RKNN要求输入为float32
+            return normalized_data.astype(np.float32)
+        
+        except Exception as e:
+            print(f"姿态数据预处理失败：{e}")
+            return None
+
+    def _get_pose_from_image(self, image_path):
+        """从图像提取姿态数据（含耗时统计+有效关键点计数）"""
+        try:
+            print(f"正在处理图像：{image_path}")
+            # 记录姿态提取开始时间
+            pose_extract_start = time.time()
+            
+            # 图像读取与预处理
+            img = Image.open(image_path).convert("RGB")
+            target_h, target_w = 257, 257  # PoseNet默认输入尺寸
+            img_resized = img.resize((target_w, target_h), Image.BILINEAR)
+            self.processed_image = np.array(img_resized, dtype=np.uint8)
+            
+            # 【关键修改1】调用PoseNet获取3个返回值：姿态数据、是否有姿态、有效关键点数量
+            pose_data, has_pose, valid_keypoint_count = get_posenet_output(image_path)
+            # 处理可能的None值（避免后续报错）
+            valid_keypoint_count = valid_keypoint_count if valid_keypoint_count is not None else 0
+            
+            # 【关键校验1】有效关键点数量不足，返回无效结果
+            if valid_keypoint_count < self.min_valid_keypoints:
+                pose_extract_time = time.time() - pose_extract_start
+                print(f"图像有效关键点不足（{valid_keypoint_count}/{self.min_valid_keypoints}），无法提取姿态")
+                return None, pose_extract_time, valid_keypoint_count
+            
+            # 姿态数据为空的情况
+            if pose_data is None or not has_pose:
+                pose_extract_time = time.time() - pose_extract_start
+                print(f"无法从图像中获取姿态数据 | 姿态提取耗时：{pose_extract_time:.4f}s")
+                return None, pose_extract_time, valid_keypoint_count
+            
+            # 姿态数据格式转换
+            pose_array = self._parse_pose_data(pose_data)
+            # 计算姿态提取耗时
+            pose_extract_time = time.time() - pose_extract_start
+            print(f"图像姿态提取完成 | 有效关键点：{valid_keypoint_count}/{self.min_valid_keypoints} | 姿态提取耗时：{pose_extract_time:.4f}s")
+            return pose_array, pose_extract_time, valid_keypoint_count
+        
+        except Exception as e:
+            pose_extract_time = time.time() - pose_extract_start
+            valid_keypoint_count = 0  # 异常时默认关键点数量为0
+            print(f"获取图像姿态数据失败：{e} | 姿态提取耗时：{pose_extract_time:.4f}s | 有效关键点：{valid_keypoint_count}")
+            return None, pose_extract_time, valid_keypoint_count
+
+    def _parse_pose_data(self, pose_data):
+        """统一解析PoseNet输出（支持字符串/数组格式）"""
+        if isinstance(pose_data, str):
+            try:
+                return np.array(json.loads(pose_data), dtype=np.float32)
+            except json.JSONDecodeError:
+                print("PoseNet输出JSON解析失败")
+                return None
+        else:
+            return np.array(pose_data, dtype=np.float32)
+
+    def _get_pose_from_frame(self, frame):
+        """从视频帧提取姿态数据（含耗时统计+有效关键点计数）"""
+        try:
+            # 记录姿态提取开始时间
+            pose_extract_start = time.time()
+            
+            # 帧格式转换：cv2默认BGR → PoseNet要求RGB
+            img_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            
+            # 尺寸调整：与PoseNet输入一致（257x257）
+            target_h, target_w = 257, 257
+            img_resized = cv2.resize(
+                img_rgb, 
+                (target_w, target_h), 
+                interpolation=cv2.INTER_LINEAR
+            )
+            processed_frame = img_resized.astype(np.uint8)  # 保留预处理后帧用于可视化
+            
+            # 【关键修改2】调用PoseNet获取3个返回值：姿态数据、是否有姿态、有效关键点数量
+            pose_data, has_pose, valid_keypoint_count = get_posenet_output(processed_frame)
+            # 处理可能的None值
+            valid_keypoint_count = valid_keypoint_count if valid_keypoint_count is not None else 0
+            
+            # 【关键校验2】有效关键点数量不足，返回无效结果
+            if valid_keypoint_count < self.min_valid_keypoints:
+                pose_extract_time = time.time() - pose_extract_start
+                print(f"帧有效关键点不足（{valid_keypoint_count}/{self.min_valid_keypoints}），无法提取姿态")
+                return None, None, pose_extract_time, valid_keypoint_count
+            
+            # 姿态数据为空的情况
+            if pose_data is None or not has_pose:
+                pose_extract_time = time.time() - pose_extract_start
+                print(f"无法从当前帧中获取姿态数据 | 姿态提取耗时：{pose_extract_time:.4f}s | 有效关键点：{valid_keypoint_count}")
+                return None, None, pose_extract_time, valid_keypoint_count
+            
+            # 统一解析姿态数据
+            pose_array = self._parse_pose_data(pose_data)
+            # 计算姿态提取耗时
+            pose_extract_time = time.time() - pose_extract_start
+            print(f"帧姿态提取完成 | 有效关键点：{valid_keypoint_count}/{self.min_valid_keypoints} | 姿态提取耗时：{pose_extract_time:.4f}s")
+            return pose_array, processed_frame, pose_extract_time, valid_keypoint_count
+        
+        except Exception as e:
+            pose_extract_time = time.time() - pose_extract_start
+            valid_keypoint_count = 0  # 异常时默认关键点数量为0
+            print(f"获取帧姿态数据失败：{e} | 姿态提取耗时：{pose_extract_time:.4f}s | 有效关键点：{valid_keypoint_count}")
+            return None, None, pose_extract_time, valid_keypoint_count
+
+    # 【新增方法1】绘制关键点不足的红色提示
+    def _draw_insufficient_keypoints(self, frame, valid_count):
+        """在帧上绘制“有效关键点不足”的红色提示文本"""
+        text = f"有效关键点不足：{valid_count}/{self.min_valid_keypoints}"
+        cv2.putText(
+            img=frame,
+            text=text,
+            org=(20, 40),  # 与正常结果文本位置一致，覆盖无效信息
+            fontFace=cv2.FONT_HERSHEY_SIMPLEX,
+            fontScale=1.0,
+            color=(0, 0, 255),  # 红色警示
+            thickness=2,
+            lineType=cv2.LINE_AA
+        )
+        # 可选：绘制FPS（即使关键点不足也显示帧率）
+        fps = 1.0 / self.last_infer_time if self.last_infer_time > 1e-6 else 0.0
+        fps_text = f"FPS: {fps:.1f}"
+        cv2.putText(
+            img=frame,
+            text=fps_text,
+            org=(20, 80),
+            fontFace=cv2.FONT_HERSHEY_SIMPLEX,
+            fontScale=0.9,
+            color=(255, 0, 0),
+            thickness=2,
+            lineType=cv2.LINE_AA
+        )
+        return frame
+
+    def _draw_result_on_frame(self, frame, formatted_result):
+        """在cv2帧上绘制推理结果"""
+        # 计算实时FPS（基于总耗时）
+        fps = 1.0 / self.last_infer_time if self.last_infer_time > 1e-6 else 0.0
+        
+        # 绘制类别+置信度
+        class_text = f"Pose: {formatted_result['class']} ({formatted_result['confidence']:.1f}%)"
+        cv2.putText(
+            img=frame,
+            text=class_text,
+            org=(20, 40),
+            fontFace=cv2.FONT_HERSHEY_SIMPLEX,
+            fontScale=1.0,
+            color=(0, 255, 0),
+            thickness=2,
+            lineType=cv2.LINE_AA
+        )
+        
+        # 绘制FPS（基于总耗时）
+        fps_text = f"FPS: {fps:.1f}"
+        cv2.putText(
+            img=frame,
+            text=fps_text,
+            org=(20, 80),
+            fontFace=cv2.FONT_HERSHEY_SIMPLEX,
+            fontScale=0.9,
+            color=(255, 0, 0),
+            thickness=2,
+            lineType=cv2.LINE_AA
+        )
+        
+        return frame
+
+    def inference_frame(self, frame, model_path=None):
+        """实时帧推理入口（含完整耗时统计+有效关键点校验）"""
+        # 记录总耗时开始时间
+        total_start = time.time()
+        result_frame = frame.copy()
+        self.processed_image = None
+        pose_data = None
+        valid_keypoint_count = 0  # 初始化有效关键点数量
+        pose_extract_time = 0.0
+
+        try:
+            # 模型加载检查
+            if model_path and (not self.rknn_lite or self.model_path != os.path.abspath(model_path)):
+                self.model_path = os.path.abspath(model_path)
+                self.load_model()
+            if not self.rknn_lite:
+                raise RuntimeError("RKNN模型未加载，无法执行推理")
+
+            # 【关键修改3】调用帧姿态提取，接收4个返回值（新增有效关键点数量）
+            pose_data, self.processed_image, pose_extract_time, valid_keypoint_count = self._get_pose_from_frame(frame)
+            
+            # 【关键校验3】有效关键点不足，直接返回无效结果
+            if valid_keypoint_count < self.min_valid_keypoints:
+                # 生成无效结果格式
+                raw_result = np.zeros(len(self.classes)) if self.classes else np.array([])
+                formatted_result = {
+                    "class": "null",
+                    "confidence": 0.0,
+                    "probabilities": raw_result.tolist(),
+                    "class_id": -1
+                }
+                # 计算总耗时并更新FPS
+                total_time = time.time() - total_start
+                self.last_infer_time = total_time
+                # 绘制关键点不足提示
+                result_frame = self._draw_insufficient_keypoints(result_frame, valid_keypoint_count)
+                # 打印日志
+                print(f"帧推理终止（有效关键点不足） | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s | FPS：{1/total_time:.1f}")
+                return raw_result, formatted_result, result_frame
+
+            # 姿态数据为空（非关键点不足的其他情况）
+            if pose_data is None or self.processed_image is None:
+                total_time = time.time() - total_start
+                self.last_infer_time = total_time
+                print(f"帧推理跳过 | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s | FPS：{1/total_time:.1f}")
+                return None, None, result_frame
+
+            # 姿态数据预处理
+            input_data = self._preprocess(pose_data)
+            if input_data is None:
+                total_time = time.time() - total_start
+                self.last_infer_time = total_time
+                print(f"帧推理失败（预处理失败） | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s | FPS：{1/total_time:.1f}")
+                return None, None, result_frame
+
+            # RKNN核心推理（单独统计推理耗时）
+            infer_start = time.time()
+            outputs = self.rknn_lite.inference(inputs=[input_data])
+            rknn_infer_time = time.time() - infer_start  # RKNN推理耗时
+            raw_output = outputs[0]
+            
+            # 结果后处理
+            raw_result = raw_output[0].flatten() if raw_output.ndim > 1 else raw_output.flatten()
+            formatted_result = self._format_result(raw_result)
+
+            # 帧结果可视化
+            result_frame = self._draw_result_on_frame(result_frame, formatted_result)
+
+            # 计算总耗时（从方法开始到推理完成）
+            total_time = time.time() - total_start
+            self.last_infer_time = total_time  # 更新最后一帧总耗时（用于FPS计算）
+            
+            # 打印完整耗时信息
+            print(f"帧推理完成 | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：{rknn_infer_time:.4f}s | FPS：{1/total_time:.1f} | 结果：{formatted_result['class']}")
+
+            return raw_result, formatted_result, result_frame
+
+        except Exception as e:
+            # 异常情况下也统计耗时
+            total_time = time.time() - total_start
+            rknn_infer_time = 0.0  # 推理未执行，耗时为0
+            self.last_infer_time = total_time
+            print(f"帧推理失败：{e} | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：{rknn_infer_time:.4f}s | FPS：{1/total_time:.1f}")
+            return None, None, result_frame
+
+    def inference(self, data, model_path=None):
+        """RKNN推理（含完整耗时统计+有效关键点校验）"""
+        # 记录总耗时开始时间
+        total_start = time.time()
+        self.processed_image = None
+        pose_data = None
+        raw_result = None
+        formatted_result = None
+        pose_extract_time = 0.0  # 默认为0（若输入已为姿态数据，无需提取）
+        valid_keypoint_count = 0  # 初始化有效关键点数量
+
+        try:
+            # 模型加载检查
+            if model_path and (not self.rknn_lite or self.model_path != os.path.abspath(model_path)):
+                self.model_path = os.path.abspath(model_path)
+                self.load_model()
+            if not self.rknn_lite:
+                raise RuntimeError("RKNN模型未加载，无法执行推理")
+            
+            # 处理输入数据
+            if isinstance(data, str):
+                # 【关键修改4】输入为图像路径：接收3个返回值（新增有效关键点数量）
+                pose_data, pose_extract_time, valid_keypoint_count = self._get_pose_from_image(data)
+                
+                # 【关键校验4】有效关键点不足，返回无效结果
+                if valid_keypoint_count < self.min_valid_keypoints:
+                    raw_result = np.zeros(len(self.classes)) if self.classes else np.array([])
+                    formatted_result = {
+                        "class": "null",
+                        "confidence": 0.0,
+                        "probabilities": raw_result.tolist(),
+                        "class_id": -1
+                    }
+                    total_time = time.time() - total_start
+                    print(f"推理终止（有效关键点不足） | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s")
+                    return raw_result, formatted_result
+                
+                # 姿态数据为空（非关键点不足的其他情况）
+                if pose_data is None:
+                    total_time = time.time() - total_start
+                    print(f"推理终止（姿态数据为空） | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s")
+                    return None, None
+            else:
+                # 输入为已提取的姿态数据，无需提取（默认关键点充足，跳过校验）
+                pose_data = data
+                valid_keypoint_count = "未知（输入为姿态数据）"
+                print(f"输入为姿态数据，跳过图像处理 | 姿态提取耗时：0.0000s | 有效关键点：{valid_keypoint_count}")
+            
+            # 数据预处理
+            input_data = self._preprocess(pose_data)
+            if input_data is None:
+                total_time = time.time() - total_start
+                print(f"推理终止（预处理失败） | 总耗时：{total_time:.4f}s | 姿态提取耗时：{pose_extract_time:.4f}s | RKNN推理耗时：0.0000s")
+                return None, None
+            
+            # RKNN推理（单独统计推理耗时）
+            infer_start = time.time()
+            outputs = self.rknn_lite.inference(inputs=[input_data])
+            rknn_infer_time = time.time() - infer_start  # RKNN推理耗时
+            raw_output = outputs[0]
+            
+            # 结果后处理
+            raw_result = raw_output[0].flatten() if raw_output.ndim > 1 else raw_output.flatten()
+            formatted_result = self._format_result(raw_result)
+            
+            # 结果可视化
+            if self.processed_image is not None and formatted_result:
+                bgr_image = cv2.cvtColor(self.processed_image, cv2.COLOR_RGB2BGR)
+                # 若为关键点不足的无效结果，绘制红色提示；否则绘制正常结果
+                if formatted_result["class"] == "null":
+                    text = f"有效关键点不足：{valid_keypoint_count}/{self.min_valid_keypoints}"
+                    cv2.putText(bgr_image, text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 0, 255), 2)
+                else:
+                    text = f"{formatted_result['class']} {formatted_result['confidence']:.1f}%"
+                    cv2.putText(bgr_image, text, (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8, (0, 255, 0), 2)
+                cv2.imwrite(self.result_image_path, bgr_image)
+                print(f"结果图已保存至：{self.result_image_path} | {text}")
+            
+            # 计算总耗时并打印完整信息
+            total_time = time.time() - total_start
+            print(f"推理完成：")
+            print(f"- 总耗时：{total_time:.4f}秒")
+            print(f"- 姿态提取耗时：{pose_extract_time:.4f}秒")
+            print(f"- RKNN推理耗时：{rknn_infer_time:.4f}秒")
+            print(f"- 有效关键点：{valid_keypoint_count}/{self.min_valid_keypoints}")
+            
+            # 将时间信息添加到formatted结果中，以便pose.py获取
+            if formatted_result:
+                formatted_result['pose_extract_time'] = pose_extract_time
+                formatted_result['inference_time'] = rknn_infer_time
+                formatted_result['total_time'] = total_time
+        
+        except Exception as e:
+            # 异常情况下也统计耗时
+            total_time = time.time() - total_start
+            rknn_infer_time = 0.0  # 推理未执行，耗时为0
+            print(f"推理失败：{e}")
+            print(f"- 总耗时：{total_time:.4f}秒")
+            print(f"- 姿态提取耗时：{pose_extract_time:.4f}秒")
+            print(f"- RKNN推理耗时：{rknn_infer_time:.4f}秒")
+            print(f"- 有效关键点：{valid_keypoint_count}/{self.min_valid_keypoints}")
+        
+        return raw_result, formatted_result
+
+    def _format_result(self, predictions):
+        """结果格式化逻辑"""
+        predictions = np.array(predictions, dtype=np.float32)
+        class_idx = np.argmax(predictions)
+        current_max = predictions[class_idx]
+        
+        # 生成格式化结果
+        confidence = float(predictions[class_idx] * 100)
+        return {
+            "class": self.classes[class_idx] if 0 <= class_idx < len(self.classes) else f"未知类别_{class_idx}",
+            "confidence": confidence,
+            "probabilities": predictions.tolist(),
+            "class_id": int(class_idx)
+        }
+
+    def show_result(self):
+        """结果显示逻辑"""
+        try:
+            result_image = cv2.imread(self.result_image_path)
+            if result_image is not None:
+                cv2.imshow("Pose RKNN Inference Result", result_image)
+                print("按任意键关闭窗口...")
+                cv2.waitKey(0)
+                cv2.destroyAllWindows()
+            else:
+                if self.processed_image is None:
+                    print("无法显示结果：输入为姿态数据，未生成图像")
+                else:
+                    print("未找到结果图片，请先执行图像路径输入的推理")
+        except Exception as e:
+            print(f"显示结果失败：{e}")
+
+    def release(self):
+        """释放RKNN资源"""
+        if hasattr(self, "rknn_lite") and self.rknn_lite:
+            self.rknn_lite.release()
+            print("RKNN NPU资源已释放")
+
+    def __del__(self):
+        """析构函数自动释放资源"""
+        self.release()
+
+    # 便捷接口
+    def predict(self, image_path):
+        """从图像路径推理（同步接口）"""
+        return self.inference(image_path)
+
+    def predict_pose_data(self, pose_data):
+        """从已提取的姿态数据推理（同步接口）"""
+        return self.inference(pose_data)
+
+
+# 实时摄像头推理测试
+if __name__ == "__main__":
+    # 配置参数（替换为你的RKNN模型路径）
+    RKNN_MODEL_PATH = "your_pose_model.rknn"
+    CAMERA_INDEX = 0  # 0=默认摄像头
+
+    # 初始化姿态推理工作流
+    pose_workflow = PoseWorkflow(model_path=RKNN_MODEL_PATH)
+    if not pose_workflow.rknn_lite:
+        print("RKNN模型初始化失败，无法启动实时推理")
+        exit(1)
+
+    # 初始化摄像头
+    cap = cv2.VideoCapture(CAMERA_INDEX)
+    if not cap.isOpened():
+        print(f"无法打开摄像头（索引：{CAMERA_INDEX}）")
+        pose_workflow.release()
+        exit(1)
+
+    # 设置摄像头分辨率
+    cap.set(cv2.CAP_PROP_FRAME_WIDTH, 640)
+    cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 480)
+    print("实时姿态推理启动成功！")
+    print(f"模型路径：{RKNN_MODEL_PATH}")
+    print(f"最小有效关键点：{pose_workflow.min_valid_keypoints}个")
+    print("按 'q' 键退出实时推理...")
+
+    # 循环读取摄像头帧并推理
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            print("无法读取摄像头帧，退出循环")
+            break
+
+        # 执行帧推理（自动打印完整耗时）
+        _, _, result_frame = pose_workflow.inference_frame(frame)
+
+        # 显示带标注的帧
+        cv2.imshow("RKNN Real-Time Pose Inference", result_frame)
+
+        # 按'q'退出
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            print("用户触发退出")
+            break
+
+    # 释放资源
+    cap.release()
+    cv2.destroyAllWindows()
+    pose_workflow.release()
+    print("实时推理结束，资源已释放")