weixin_41875016的博客

Mediapipe posedetection 源码分析正文Pose classifierPoseClassifierProcessor.java    作用：Accepts a stream of Pose for classification and Rep counting.接受Pose流来进行分类和Rep计数    主要参数：POSE_SAMPLES_FILE = "pose/fitness_pose_samples.csv"储存pose示例的csv文件;    String[]

Mediapipe Blazepose poseclassification android 源码分析（二）——ClassificationResult.javaClassificationResult.java    作用：表示Posecasifier输出的Pose分类结果。可以修改。    变量：Map<String, Float> classConfidences只有这一个变量，    其中key是className，    value是classConfidences，含义

EMASmoothing.java    作用：Runs EMA smoothing over a window with given stream of pose classification results.    意思是：在给定姿势分类结果的流的窗口上运行EMA平滑处理    基本参数：DEFAULT_WINDOW_SIZE = 10窗口大小默认为10；    DEFAULT_ALPHA = 0.2f默认α为0.2，代表了新旧结果所占的权重，权重衰减速度α成正比；    window是一.

RepetitionCounter.java    统计动作次数的关键代码    作用：Counts reps for the give class.为给定的类统计重复次数    参数：className，enterThreshold = 6f，exitThreshold = 4f，numRepeats = 0，poseEntered = false    实例化方法：就是初始化各个参数    重要方法：    - addClassificationResult(Classification.

PoseEmbedding.java    作用：Generates embedding for given list of Pose landmarks.    参数：TORSO_MULTIPLIER = 2.5f适用于躯干，以获得最小的身体尺寸的乘数。通过实验选择了这个。    - 方法：        - getPoseSize(List<PointF3D> landmarks)            详细步骤大致是：先求出hipsCenter和shouldersCenter.

PoseClassifier.java    作用：Classifies Pose based on given PoseSamples. 基于pose的示例给pose分类。    思想：KNN    主要参数：MAX_DISTANCE_TOP_K = 30；MEAN_DISTANCE_TOP_K = 10;    AXES_WEIGHTS = PointF3D.from(1, 1, 0.2f)因为Z轴相对于X和Y对精度影响小，所以所占权重默认0.2；    所有参数均可人为设置    - .