import cv2 import mediapipe as mp import math import pandas as pd import numpy as np from sklearn.model_selection import train_test_split from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout

初始化MediaPipe的人体姿势模型

mp_drawing = mp.solutions.drawing_utils mp_pose = mp.solutions.pose

打开输入视频文件

cap = cv2.VideoCapture('6.mp4')

获取输入视频的帧率和分辨率

fps = int(cap.get(cv2.CAP_PROP_FPS)) width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))

创建输出视频文件

fourcc = cv2.VideoWriter_fourcc(*'mp4v') out = cv2.VideoWriter('9_1.mp4', fourcc, fps, (width, height))

加载动作数据集

warmup_df = pd.read_csv('dataset_warm-up.csv') combat_df = pd.read_csv('dataset_SHIZHAN POSE.csv') attack_df = pd.read_csv('dataset_hit.csv') respect_df = pd.read_csv('dataset_respect.csv')

将数据集合并为一个大的数据集

data = pd.concat([warmup_df, combat_df, attack_df, respect_df], ignore_index=True)

将数据集拆分为训练集和测试集

X_train, X_test, y_train, y_test = train_test_split( data[['angle1','angle2','angle3', 'angle4','angle5' ,'angle6', "angle7","angle8","angle9","angle10","angle11"]], data['label'], test_size=0.2)

将数据集转换为卷积神经网络的输入格式

X_train = np.array(X_train).reshape(-1, 11, 1, 1) X_test = np.array(X_test).reshape(-1, 11, 1, 1)

构建卷积神经网络模型

model = Sequential([ Conv2D(32, (3, 1), activation='relu', input_shape=(11, 1, 1)), MaxPooling2D((2, 1)), Conv2D(64, (3, 1), activation='relu'), MaxPooling2D((2, 1)), Conv2D(128, (3, 1), activation='relu'), MaxPooling2D((2, 1)), Flatten(), Dense(128, activation='relu'), Dropout(0.5), Dense(4, activation='softmax') ])

编译模型

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

训练模型

model.fit(X_train, y_train, epochs=10, validation_data=(X_test, y_test))

处理视频文件中的每一帧

with mp_pose.Pose(min_detection_confidence=0.5, min_tracking_confidence=0.5) as pose: while cap.isOpened(): # 读取一帧 ret, frame = cap.read() if not ret: break

    # 将帧转换为RGB格式
    image = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)

    # 处理人体姿势检测
    results = pose.process(image)

    # 判断是否检测到人体
    if results.pose_landmarks:
        # 绘制人体骨架
        mp_drawing.draw_landmarks(
            frame, results.pose_landmarks, mp_pose.POSE_CONNECTIONS)

        # 获取鼻子、右肩和右手腕关键点的信息
        nose = results.pose_landmarks.landmark[mp_pose.PoseLandmark.NOSE]
        right_shoulder = results.pose_landmarks.landmark[mp_pose.PoseLandmark.RIGHT_SHOULDER]
        right_wrist = results.pose_landmarks.landmark[mp_pose.PoseLandmark.RIGHT_WRIST]

        # 获取鼻子、右肩和右手腕的角度
        angle_nwr = math.degrees(math.atan2(right_wrist.y - nose.y, right_wrist.x - nose.x) -
                                 math.atan2(right_shoulder.y - nose.y, right_shoulder.x - nose.x))

        # 将数据输入卷积神经网络进行预测
        X = np.array([[angle, angle1, angle_dl, angle_dr, angle_tr, angle_tl,angle_lka,angle_hls,angle_rka,angle_hrs,angle_nwr]])
        X = X.reshape(-1, 11, 1, 1)
        label = model.predict(X)
        label = np.argmax(label)

        # 在输出图片上显示角度值和动作类型
        cv2.putText(frame, "Angle: {:.2f}".format(angle), (5, 43),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 2)
        cv2.putText(frame, label, (600, 450),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (100, 225, 225), 2)
        cv2.putText(frame, label, (5, 20),
                    cv2.FONT_HERSHEY_COMPLEX_SMALL, 1.5, (225, 100, 225), 2)

    else:
        # 如果未检测到人体，则跳过本帧处理
        cv2.putText(frame, "No body detected", (50, 50),
                    cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)

    # 将帧写入输出视频文件
    out.write(frame)

    # 显示当前帧的结果
    cv2.imshow('MediaPipe Pose Detection press q exit', frame)

    # 检测是否按下q键退出
    if cv2.waitKey(1) & 0xFF == ord('q'):
        break

释放资源

cap.release() out.release() cv2.destroyAllWindows()