mediaplayer能播放什么格式（Mediapipe实战——导出身体节点坐标并用TensorFlow搭建LSTM网络来训练自己的手势检测模型再部署到树莓派4B）

时间2025-09-16 16:11:13分类IT科技浏览6357

导读：一、前言...

一、前言

在YouTube上看到up主——Nicholas Renotte的相关教程，觉得非常有用。使用他的方法，我训练了能够检测四种手势的模型，在这里和大家分享一下。

附上该up主的视频链接Sign Language Detection using ACTION RECOGNITION with Python | LSTM Deep Learning Model

视频的代码链接https://github.com/nicknochnack/ActionDetectionforSignLanguage

我的系列文章一：Mediapipe入门——搭建姿态检测模型并实时输出人体关节点3d坐标

我的系列文章二：Mediapipe姿态估计——用坐标计算手指关节弯曲角度并实时标注

我使用的环境

Pycharm2021

mediapipe0.8.9

tensorflow2.3.0

openCV4.5.4

个人认为版本影响不大，可以跟我不一致，但tensorflow最好2.0以上

二、使用mediapipe搭建姿态估计模型并打开摄像头采集坐标数据集

源代码中，up主进行了很好地封装，代码稍长，接下来我只挑重要的部分说一下，完整的代码请看文末（代码中的中文注释是我添加的，英文的是原作者的）。

首先是处理视频流的函数。 def mediapipe_detection(image, model): image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # BGR 转 RGB image.flags.writeable = False # Image is no longer writeable results = model.process(image) # 对视频流处理，返回坐标 image.flags.writeable = True # Image is now writeable image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) # RGB 转 BGR return image, results

然后是在人体上渲染节点的函数。

def draw_styled_landmarks(image, results): mp_drawing.draw_landmarks(image, results.pose_landmarks, mp_holistic.POSE_CONNECTIONS, mp_drawing.DrawingSpec(color=(80,22,10), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(80,44,121), thickness=2, circle_radius=2) ) ...... #剩下还有，不一一放上来了，完整请看文末

这两个功能比较简单，如果想了解如何用mediapipe搭建姿态检测模型，请看我的系列文章一。

然后是比较重要的提取坐标的函数，将process返回的坐标提取出来，并转换为numpy矩阵。为了训练手势模型，我使用了姿势坐标33个、左右手坐标各21个。原作者还使用了脸部坐标一起训练，个人没这个需求，将相关代码注释了。 def extract_keypoints(results): #姿势坐标33个，np.zeros(33*4)是因为除x,y,z外，还有置信度visibility ，以下类似 pose = np.array([[res.x, res.y, res.z, res.visibility] for res in results.pose_landmarks.landmark]).flatten() if results.pose_landmarks else np.zeros(33*4) #mediapipe面网多达468个节点，这里我不用，注释掉 #face = np.array([[res.x, res.y, res.z] for res in results.face_landmarks.landmark]).flatten() if results.face_landmarks else np.zeros(468*3) #左手坐标21个 lh = np.array([[res.x, res.y, res.z] for res in results.left_hand_landmarks.landmark]).flatten() if results.left_hand_landmarks else np.zeros(21*3) #右手坐标21个 rh = np.array([[res.x, res.y, res.z] for res in results.right_hand_landmarks.landmark]).flatten() if results.right_hand_landmarks else np.zeros(21*3) return np.concatenate([pose, lh, rh]) #如果要使用脸部坐标训练，列表更换为[pose, face, lh, rh]

33个姿势节点如下所示。

21个手部节点如下所示。

现在使用os库在同一目录下新建文件夹存放等下要采集的数据集。

DATA_PATH = os.path.join(MP_Data)

接下来比较重要了。我将训练的四个手势是“666 ” ，“大拇指 ” 、“比心 ” 、“剪刀手 ” 。每个动作将采集30次，每次采集30帧（这些可以改）

actions = np.array([666, thumbs_up, finger_heart,scissor_hand])#你要训练的手势名称，即动作标签label # Thirty videos worth of data no_sequences = 30#采集30次 # Videos are going to be 30 frames in length sequence_length = 30#30帧 #关于这个for循环，会在MP_data文件下建立四个文件夹（对应四个动作），每个文件夹又包含30个子文件夹， #每个子文件夹包含30个.npy文件，都是每次采集坐标信息时保存的 for action in actions: for sequence in range(no_sequences): try: os.makedirs(os.path.join(DATA_PATH, action, str(sequence))) except: pass

然后运行这部分程序开始采集数据集（完整代码请看文末）。采集前都会有提示，原作者做得很好。

就这样慢慢采集，大概几分钟，采集完会自动结束程序。

三、使用Tensorflow搭建LSTM网络进行训练，然后保存模型有了数据集，开始搭建网络训练。关于长短期记忆网络LSTM ，请看官网的介绍 #同样，这里只是部分代码，详细请看文末 from tensorflow.keras.models import Sequential from tensorflow.keras.layers import LSTM, Dense model = Sequential() #关于input_shape ，原作者的网络是(30,1662),1662=33*4 + 468*2 + 21*3 + 21*3 ，而我不需要面网坐标，故只有258 model.add(LSTM(64, return_sequences=True, activation=relu, input_shape=(30,258))) model.add(LSTM(128, return_sequences=True, activation=relu)) model.add(LSTM(64, return_sequences=False, activation=relu)) model.add(Dense(64, activation=relu)) model.add(Dense(32, activation=relu)) model.add(Dense(actions.shape[0], activation=softmax)) model.compile(optimizer=Adam, loss=categorical_crossentropy, metrics=[categorical_accuracy]) model.fit(X_train, y_train, epochs=2000, callbacks=[tb_callback]) model.summary() model.save(action.h5)#要保存的模型名称，保存在当前目录

tensorflow的使用还是比较简单的，如果看不懂，请看TensorFlow中文官网。训练结果如图。

虽有2000个epochs，但即使是CPU下训练速度也很快。最后在同一目录下得到了我们的权重文件action.h5 ，接下来就可以实际使用训练好的模型了。

四、使用训练好的模型进行实际检测

看效果图吧，当识别到对应手势，相应标签的框框颜色条会变长，这代表分类到这一手势的概率。同时运行端也会输出此刻检测到手势类别。这部分代码与上文的代码大体类似，请看文末吧。总的来说，手势基本上都能识别正确，响应速度也很快。

最后我将该模型部署到了树莓派上，虽然运行起来有点慢，但还是很成功的。部署的话，就是注意相关库都要安装，然后代码和权重文件拖过去运行就好了，没什么难点。

五、总结 借助该up主的代码，可以简便的训练自己的手势识别模型，准确率也高。不过要注意的是，当使用训练好的模型进行实际检测时，所做动作务必和采集数据集时的动作保持一致 。这是因为，代码中使用的mediapipe坐标会随你离摄像头的距离变化而变化。所以同样的手势动作，只要你离摄像头的距离或角度变了，识别准确率就会大大下降，这是我多次实践得出的结论。使用自己的模型时，所做动作务必和采集数据集时的动作保持一致！

六、所有代码

如果你想复现我的模型，你不需要改动任何代码；如果想扩大数据集，请修改no_sequences 和sequence_length；如果想训练别的动作或增加动作数目，请修改actions列表和colors列表（增加或减少动作数目就要修改）；想训练面网坐标，增加表情识别，请取消相应注释。如果有其他不懂的，可以在评论区问我。

首先是采集数据集的代码 import cv2 import numpy as np import os import mediapipe as mp mp_holistic = mp.solutions.holistic # Holistic model mp_drawing = mp.solutions.drawing_utils # Drawing utilities def mediapipe_detection(image, model): image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # COLOR CONVERSION BGR 2 RGB image.flags.writeable = False # Image is no longer writeable results = model.process(image) # Make prediction image.flags.writeable = True # Image is now writeable image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) # COLOR COVERSION RGB 2 BGR return image, results def draw_styled_landmarks(image, results): """ 要训练脸部坐标就取消注释 # Draw face connections mp_drawing.draw_landmarks(image, results.face_landmarks, mp_holistic.FACEMESH_CONTOURS, mp_drawing.DrawingSpec(color=(80,110,10), thickness=1, circle_radius=1), mp_drawing.DrawingSpec(color=(80,256,121), thickness=1, circle_radius=1) ) """ # Draw pose connections mp_drawing.draw_landmarks(image, results.pose_landmarks, mp_holistic.POSE_CONNECTIONS, mp_drawing.DrawingSpec(color=(80,22,10), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(80,44,121), thickness=2, circle_radius=2) ) # Draw left hand connections mp_drawing.draw_landmarks(image, results.left_hand_landmarks, mp_holistic.HAND_CONNECTIONS, mp_drawing.DrawingSpec(color=(121,22,76), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(121,44,250), thickness=2, circle_radius=2) ) # Draw right hand connections mp_drawing.draw_landmarks(image, results.right_hand_landmarks, mp_holistic.HAND_CONNECTIONS, mp_drawing.DrawingSpec(color=(245,117,66), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(245,66,230), thickness=2, circle_radius=2) ) def extract_keypoints(results): pose = np.array([[res.x, res.y, res.z, res.visibility] for res in results.pose_landmarks.landmark]).flatten() if results.pose_landmarks else np.zeros(33*4) #face = np.array([[res.x, res.y, res.z] for res in results.face_landmarks.landmark]).flatten() if results.face_landmarks else np.zeros(468*3) lh = np.array([[res.x, res.y, res.z] for res in results.left_hand_landmarks.landmark]).flatten() if results.left_hand_landmarks else np.zeros(21*3) rh = np.array([[res.x, res.y, res.z] for res in results.right_hand_landmarks.landmark]).flatten() if results.right_hand_landmarks else np.zeros(21*3) return np.concatenate([pose, lh, rh]) # Path for exported data, numpy arrays DATA_PATH = os.path.join(MP_Data) # Actions that we try to detect actions = np.array([666, thumbs_up, finger_heart,scissor_hand]) # Thirty videos worth of data no_sequences = 30 # Videos are going to be 30 frames in length sequence_length = 30 for action in actions: for sequence in range(no_sequences): try: os.makedirs(os.path.join(DATA_PATH, action, str(sequence))) except: pass cap = cv2.VideoCapture(0) # Set mediapipe model with mp_holistic.Holistic(min_detection_confidence=0.5, min_tracking_confidence=0.5) as holistic: # NEW LOOP # Loop through actions for action in actions: # Loop through sequences aka videos for sequence in range(no_sequences): # Loop through video length aka sequence length for frame_num in range(sequence_length): # Read feed ret, frame = cap.read() # Make detections image, results = mediapipe_detection(frame, holistic) # print(results) # Draw landmarks draw_styled_landmarks(image, results) # NEW Apply wait logic if frame_num == 0: cv2.putText(image, STARTING COLLECTION, (120, 200), cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 4, cv2.LINE_AA) cv2.putText(image, Collecting frames for {} Video Number {}.format(action, sequence), (15, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA) # Show to screen cv2.imshow(OpenCV Feed, image) cv2.waitKey(2000) else: cv2.putText(image, Collecting frames for {} Video Number {}.format(action, sequence), (15, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA) # Show to screen cv2.imshow(OpenCV Feed, image) # NEW Export keypoints keypoints = extract_keypoints(results) npy_path = os.path.join(DATA_PATH, action, str(sequence), str(frame_num)) np.save(npy_path, keypoints) # Break gracefully if cv2.waitKey(10) & 0xFF == ord(q): break cap.release() cv2.destroyAllWindows()

使用TensorFlow搭建LSTM网络进行训练

from tensorflow.keras.models import Sequential from tensorflow.keras.layers import LSTM, Dense from tensorflow.keras.callbacks import TensorBoard import numpy as np import os from sklearn.model_selection import train_test_split from tensorflow.keras.utils import to_categorical log_dir = os.path.join(Logs) tb_callback = TensorBoard(log_dir=log_dir) no_sequences = 30 # Videos are going to be 30 frames in length sequence_length = 30 DATA_PATH = os.path.join(MP_Data) actions = np.array([666, thumbs_up, finger_heart,scissor_hand]) label_map = {label:num for num, label in enumerate(actions)} sequences, labels = [], [] for action in actions: for sequence in range(no_sequences): window = [] for frame_num in range(sequence_length): res = np.load(os.path.join(DATA_PATH, action, str(sequence), "{}.npy".format(frame_num))) window.append(res) sequences.append(window) labels.append(label_map[action]) X = np.array(sequences) y = to_categorical(labels).astype(int) X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.05) model = Sequential() model.add(LSTM(64, return_sequences=True, activation=relu, input_shape=(30,258))) model.add(LSTM(128, return_sequences=True, activation=relu)) model.add(LSTM(64, return_sequences=False, activation=relu)) model.add(Dense(64, activation=relu)) model.add(Dense(32, activation=relu)) model.add(Dense(actions.shape[0], activation=softmax)) model.compile(optimizer=Adam, loss=categorical_crossentropy, metrics=[categorical_accuracy]) model.fit(X_train, y_train, epochs=2000, callbacks=[tb_callback]) model.summary() model.save(action.h5)

使用训练好的模型进行实际检测

from tensorflow.keras.models import Sequential from tensorflow.keras.layers import LSTM, Dense import cv2 import numpy as np import mediapipe as mp mp_holistic = mp.solutions.holistic # Holistic model mp_drawing = mp.solutions.drawing_utils # Drawing utilities sequence = [] sentence = [] threshold = 0.8 actions = np.array([666, thumbs_up, finger_heart,scissor_hand]) def mediapipe_detection(image, model): image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # COLOR CONVERSION BGR 2 RGB image.flags.writeable = False # Image is no longer writeable results = model.process(image) # Make prediction image.flags.writeable = True # Image is now writeable image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) # COLOR COVERSION RGB 2 BGR return image, results def draw_styled_landmarks(image, results): # Draw face connections """ mp_drawing.draw_landmarks(image, results.face_landmarks, mp_holistic.FACEMESH_CONTOURS, mp_drawing.DrawingSpec(color=(80,110,10), thickness=1, circle_radius=1), mp_drawing.DrawingSpec(color=(80,256,121), thickness=1, circle_radius=1) ) """ # Draw pose connections mp_drawing.draw_landmarks(image, results.pose_landmarks, mp_holistic.POSE_CONNECTIONS, mp_drawing.DrawingSpec(color=(80,22,10), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(80,44,121), thickness=2, circle_radius=2) ) # Draw left hand connections mp_drawing.draw_landmarks(image, results.left_hand_landmarks, mp_holistic.HAND_CONNECTIONS, mp_drawing.DrawingSpec(color=(121,22,76), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(121,44,250), thickness=2, circle_radius=2) ) # Draw right hand connections mp_drawing.draw_landmarks(image, results.right_hand_landmarks, mp_holistic.HAND_CONNECTIONS, mp_drawing.DrawingSpec(color=(245,117,66), thickness=2, circle_radius=4), mp_drawing.DrawingSpec(color=(245,66,230), thickness=2, circle_radius=2) ) def extract_keypoints(results): pose = np.array([[res.x, res.y, res.z, res.visibility] for res in results.pose_landmarks.landmark]).flatten() if results.pose_landmarks else np.zeros(33*4) #face = np.array([[res.x, res.y, res.z] for res in results.face_landmarks.landmark]).flatten() if results.face_landmarks else np.zeros(468*3) lh = np.array([[res.x, res.y, res.z] for res in results.left_hand_landmarks.landmark]).flatten() if results.left_hand_landmarks else np.zeros(21*3) rh = np.array([[res.x, res.y, res.z] for res in results.right_hand_landmarks.landmark]).flatten() if results.right_hand_landmarks else np.zeros(21*3) return np.concatenate([pose, lh, rh]) model = Sequential() model.add(LSTM(64, return_sequences=True, activation=relu, input_shape=(30,258))) model.add(LSTM(128, return_sequences=True, activation=relu)) model.add(LSTM(64, return_sequences=False, activation=relu)) model.add(Dense(64, activation=relu)) model.add(Dense(32, activation=relu)) model.add(Dense(actions.shape[0], activation=softmax)) model.load_weights(action.h5) colors = [(245, 117, 16), (117, 245, 16), (16, 117, 245),(16, 117, 245)]#四个动作的框框，要增加动作数目，就多加RGB元组 def prob_viz(res, actions, input_frame, colors): output_frame = input_frame.copy() for num, prob in enumerate(res): cv2.rectangle(output_frame, (0, 60 + num * 40), (int(prob * 100), 90 + num * 40), colors[num], -1) cv2.putText(output_frame, actions[num], (0, 85 + num * 40), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2, cv2.LINE_AA) return output_frame cap = cv2.VideoCapture(0) # Set mediapipe model with mp_holistic.Holistic(min_detection_confidence=0.5, min_tracking_confidence=0.5) as holistic: while cap.isOpened(): # Read feed ret, frame = cap.read() # Make detections image, results = mediapipe_detection(frame, holistic) print(results) # Draw landmarks draw_styled_landmarks(image, results) # 2. Prediction logic keypoints = extract_keypoints(results) sequence.append(keypoints) sequence = sequence[-30:] if len(sequence) == 30: res = model.predict(np.expand_dims(sequence, axis=0))[0] print(actions[np.argmax(res)]) # 3. Viz logic if res[np.argmax(res)] > threshold: if len(sentence) > 0: if actions[np.argmax(res)] != sentence[-1]: sentence.append(actions[np.argmax(res)]) else: sentence.append(actions[np.argmax(res)]) if len(sentence) > 5: sentence = sentence[-5:] # Viz probabilities image = prob_viz(res, actions, image, colors) cv2.rectangle(image, (0, 0), (640, 40), (245, 117, 16), -1) cv2.putText(image, .join(sentence), (3, 30), cv2.FONT_HERSHEY_SIMPLEX, 1, (255, 255, 255), 2, cv2.LINE_AA) # Show to screen cv2.imshow(OpenCV Feed, image) # Break gracefully if cv2.waitKey(10) & 0xFF == ord(q): break cap.release() cv2.destroyAllWindows()

七、我也只是搬运工，欢迎在评论区讨论、赐教