MoveNet实时摄像头姿态检测关键点定位不准问题求助

• 在搭载M2芯片的13英寸MacBook Pro上，使用MoveNet处理摄像头实时画面时出现以下定位偏差：
  • 仅拍摄人脸及肩部上方区域时，肩部关键点位置过高
  • 受试者后移至画面完整显示上半身时，肩部关键点恢复正常，但眼部位置偏低，且手臂关键点仅能识别到肘部，无法延伸至手腕
• 尝试修改绘图与预处理函数后，有时会出现关键点直接偏移至屏幕左上角，完全无法对应人体区域的情况

import numpy as np
from matplotlib import pyplot as plt
import cv2 
import tensorflow as tf
EDGES = {
    (0, 1): 'm',
    (0, 2): 'c',
    (1, 3): 'm',
    (2, 4): 'c',
    (0, 5): 'm',
    (0, 6): 'c',
    (5, 7): 'm',
    (7, 9): 'm',
    (6, 8): 'c',
    (8, 10): 'c',
    (5, 6): 'y',
    (5, 11): 'm',
    (6, 12): 'c',
    (11, 12): 'y',
    (11, 13): 'm',
    (13, 15): 'm',
    (12, 14): 'c',
    (14, 16): 'c'
}# 定义关键点之间的连接边


def draw_keypoints(frame, keypoints, confidence_threshold):
    y, x, c = frame.shape
    shaped = np.squeeze(np.multiply(keypoints, [y,x,1]))
    
    for kp in shaped:
        ky, kx, kp_conf = kp
        if kp_conf > confidence_threshold:
            cv2.circle(frame, (int(kx), int(ky)), 4, (0,255,0), -1) 

def draw_connections(frame, keypoints, edges, confidence_threshold):
    y, x, c = frame.shape
    shaped = np.squeeze(np.multiply(keypoints, [y,x,1]))
    
    for edge, color in edges.items():
        p1, p2 = edge
        y1, x1, c1 = shaped[p1]
        y2, x2, c2 = shaped[p2]
        
        if (c1 > confidence_threshold) & (c2 > confidence_threshold):      
            cv2.line(frame, (int(x1), int(y1)), (int(x2), int(y2)), (0,0,255), 2)

def preprocess_image(frame):
    # 定义目标尺寸
    target_size = 256

    # 计算原始画面的宽高比
    orig_height, orig_width, _ = frame.shape
    aspect_ratio = orig_width / orig_height

    # 调整画面尺寸
    if aspect_ratio >= 1:  # 宽大于等于高
        new_width = target_size
        new_height = round(target_size / aspect_ratio)
    else:  # 高大于宽
        new_height = target_size
        new_width = round(target_size * aspect_ratio)
    frame = cv2.resize(frame, (new_width, new_height))

    # 填充画面
    pad_top = (target_size - new_height) // 2
    pad_bottom = target_size - new_height - pad_top
    pad_left = (target_size - new_width) // 2
    pad_right = target_size - new_width - pad_left
    frame = cv2.copyMakeBorder(frame, pad_top, pad_bottom, pad_left, pad_right, cv2.BORDER_CONSTANT)

    return frame


interpreter = tf.lite.Interpreter(model_path='lite-model_movenet_singlepose_thunder_3.tflite') # 加载模型
interpreter.allocate_tensors() # 为模型分配内存
img = any
cap = cv2.VideoCapture(0)
while cap.isOpened():
    ret, frame = cap.read()
    
    # 重塑图像
    img = frame.copy()
    img = preprocess_image(img)
    # 转换为float32并添加batch维度
    input_image = np.expand_dims(img.astype(np.float32), axis=0)

    
    # 设置输入输出
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    
    # 执行预测
    interpreter.set_tensor(input_details[0]['index'], input_image)
    interpreter.invoke()
    keypoints_with_scores = interpreter.get_tensor(output_details[0]['index'])
    
    # 渲染关键点和连接边
    draw_connections(frame, keypoints_with_scores, EDGES, 0.1)
    draw_keypoints(frame, keypoints_with_scores, 0.1)
    
    cv2.imshow('MoveNet Lightning', frame)
    
    if cv2.waitKey(10) & 0xFF==ord('q'):
        break
        
cap.release()
cv2.destroyAllWindows()
plt.imshow(img) # 显示预处理后的图像
print(img.shape)

right_hand = keypoints_with_scores[0][0][9] # 获取右手关键点
left_hand = keypoints_with_scores[0][0][10] # 获取左手关键点
px_cordinates = np.array(left_hand[:2]*[720,1280]).astype(int) # 转换为像素坐标

内容的提问来源于stack exchange，提问作者Agam Aneja