【人工智能学习之姿态估计】

算法思路

通过模型可以得到鼻子，左眼，右眼，左耳，右耳，左肩，右肩，左肘，右肘，左腕，右腕，左臀，右臀，左膝，右膝，左踝，右踝的17个关节点，通过这些关节点设计算法即可判断人体姿态。

姿态估计可以直接使用训练好的模型和权重model=“yolo_nas_pose_l”, weights=“coco_pose”。

算法实现

import cv2
import torch
from PIL import Image
from torchvision import transforms
import super_gradients
from super_gradients.training import models
from super_gradients.common.object_names import Models
import math
import matplotlib
matplotlib.use('TkAgg')
point_name = ['鼻子','左眼','右眼','左耳','右耳','左肩','右肩','左肘','右肘','左腕','右腕','左臀','右臀','左膝','右膝','左踝','右踝']
keypoints = [
    "nose",
    "left_eye",
    "right_eye",
    "left_ear",
    "right_ear",
    "left_shoulder",
    "right_shoulder",
    "left_elbow",
    "right_elbow",
    "left_wrist",
    "right_wrist",
    "left_hip",
    "right_hip",
    "left_knee",
    "right_knee",
    "left_ankle",
    "right_ankle"
]
transform = transforms.Compose([
    transforms.Resize((640, 640)),
    transforms.ToTensor(),
])

def image_to_tensor(image_path):
    img = Image.open(image_path).convert('RGB')
    tensor = transform(img)
    tensor = tensor.unsqueeze(0)
    return tensor

def pose_piont_txt(poses,creen):
    for idx,num_joints in enumerate(poses):
        x, y, confidence = num_joints
        cv2.putText(creen,f'{
      
      keypoints[idx]}',(int(x),int(y)),fontFace=cv2.FONT_HERSHEY_SIMPLEX ,fontScale=1,color=(255,0,0),thickness=1)


def posture_detection(poses):
    nose, left_eye, right_eye, left_ear, right_ear, left_shoulder, right_shoulder, \
    left_elbow, right_elbow, left_wrist, right_wrist, left_hip, right_hip, \
    left_knee, right_knee, left_ankle, right_ankle = poses

    mid_hip = ((left_hip[0] + right_hip[0]) / 2, (left_hip[1] + right_hip[1]) / 2)

    # 脊柱角度
    spine_angle = math.degrees(math.atan2(mid_hip[1] - nose[1], mid_hip[0] - nose[0]))

    # 躯干角度
    torso_angle = math.degrees(math.atan2(right_hip[1] - right_shoulder[1], right_hip[0] - right_shoulder[0]))

    # 左右腿角度
    left_leg_angle = math.degrees(math.atan2(left_ankle[1] - left_knee[1], left_ankle[0] - left_knee[0]))
    right_leg_angle = math.degrees(math.atan2(right_ankle[1] - right_knee[1], right_ankle[0] - right_knee[0]))

    # 肩部与髋部的相对高度
    shoulder_hip_diff = (left_shoulder[1] + right_shoulder[1]) / 2 - (left_hip[1] + right_hip[1]) / 2

    # 判断姿势
    if abs(spine_angle) > 75 and abs(torso_angle) > 75:
        return "Lying Down"
    elif abs(shoulder_hip_diff) < 50 and abs(torso_angle) < 30 and min(left_leg_angle, right_leg_angle) < 150:
        return "Sitting"
    elif abs(shoulder_hip_diff) > 100 and abs(torso_angle) < 30 and max(left_leg_angle, right_leg_angle) > 160:
        return "Standing"
    else:
        return "Standing"


def detect_img():
    image_path0 = 'data/test0.jpeg'
    image_path1 = 'data/test1.jpeg'
    image_path2 = 'data/test2.jpg'
    # tensor_image = image_to_tensor(image_path)
    # print(tensor_image.shape)

    # net = torch.jit.load("yolo_nas_m.pt")
    # output = net(tensor_image)
    # print(output)


    yolo_nas = super_gradients.training.models.get("yolo_nas_pose_l", pretrained_weights="coco_pose").cuda()
    frame = cv2.imread(image_path2)
    img = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    txt_img = frame.copy()
    model_predictions = yolo_nas.predict(img, conf=0.8)
    # pose_image = model_predictions.image
    prediction = model_predictions.prediction
    bboxes = prediction.bboxes_xyxy
    poses = prediction.poses
    scores = prediction.scores
    for idx in range(len(bboxes)):
        x1 = int(bboxes[idx][0])
        y1 = int(bboxes[idx][1])
        x2 = int(bboxes[idx][2])
        y2 = int(bboxes[idx][3])
        pose_piont_txt(poses[idx], txt_img)
        posture = posture_detection(poses[idx])
        creen = txt_img[y1:y2,x1:x2,:]
        cv2.putText(img, f'{
      
      posture}', (x1+60, y1), fontFace=cv2.FONT_HERSHEY_SIMPLEX, fontScale=1, color=(0, 0, 255), thickness=2)
        cv2.imshow(f'creen{
      
      idx}',creen)

    model_predictions.show()
    cv2.waitKey(0)

def detact_vidio():
    vidio_path = 'data/cs4.mp4'
    yolo_nas = super_gradients.training.models.get("yolo_nas_pose_l", pretrained_weights="coco_pose").cuda()
    model_predictions = yolo_nas.predict(vidio_path, conf=0.8)
    model_predictions.save('data/cs5.mp4')

if __name__ == '__main__':
    detect_img()
    # detact_vidio()

效果图