python -m tf2onnx.convert --opset 11 --tflite pose_landmark_lite.tflite --output pose_landmark_lite.onnx

onnx 推理代码：

import os
import cv2
import onnxruntime
import time
os.environ['TF_CPP_MIN_LOG_LEVEL']='3'
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
import numpy as np
if __name__ == '__main__':
    for aa in [1,2]:
        # img_path=f'./166_{aa}.jpg'
        img_path=f'./03.jpg'
        img=cv2.imread(img_path)
        if img is None:
            exit(1)
        t_h, t_w = img.shape[:2]
        a_w = max(t_w, t_h)
        img_b = np.zeros((a_w, a_w, 3), dtype=np.uint8)
        if t_h > t_w:
            img_b[:, 

AI识别人可以分成五个层次，依次为： 1.有没有人？ object detection 2.人在哪里？ object localization & semantic segmentation 3.这个人是谁？ face identification 4.这个人此刻处于什么状态？ pose estimation 5.这个人在当前一段时间里在做什么？ Sequence action recognition # Set up the model MODEL_WEIGHTS = 'resnet18_baseline_att_224x224_A_epoch_249.pth' OPTIMIZED_MODEL = 'resnet18_baseline_att_224x224_A_epoch_249_trt.pth' WIDTH = 224 HEIGHT = 224 # Load the model model = trt_ pose .models.resnet18_baseline_att(num_classes=2, pretrained=True) model.load_state_dict(torch.load(MODEL_WEIGHTS)) model = model.cuda().eval() # Set up the optimized model data = torch.zeros((1, 3, HEIGHT, WIDTH)).cuda() model_trt = torch2trt.torch2trt(model, [data], fp16_mode=True, max_workspace_size=1<<25) torch.save(model_trt.state_dict(), OPTIMIZED_MODEL) # Load the optimized model model_trt = TRTModule() model_trt.load_state_dict(torch.load(OPTIMIZED_MODEL)) # Load the COCO classes with o pe n('human_ pose .json', 'r') as f: human_ pose = json.load(f) # Load the keypoints num_parts = len(human_ pose ['keypoints']) part_names = [] part_edges = [] for i in range(num_parts): part_names.ap pe nd(human_ pose ['keypoints'][i]) if 'skeleton' in human_ pose : skel = human_ pose ['skeleton'] for j in range(len(skel)): if skel[j][0] == i: part_edges.ap pe nd((skel[j][0], skel[j][1])) # Set up the transforms transform = transforms.Com pose ([ transforms.Resize((HEIGHT, WIDTH)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # Set up the camera cap = cv2.VideoCapture(0) if not cap.isO pe ned(): print("Cannot o pe n camera") exit() while True: # Capture frame-by-frame ret, frame = cap.read() # Process the frame image = PIL.Image.fromarray(frame).convert('RGB') image = transform(image).cuda() cmap, paf = model_trt(image[None, ...]) cmap, paf = cmap.detach().cpu(), paf.detach().cpu() counts, objects, pe aks = trt_ pose .models.parse_objects(cmap, paf) # Draw the keypoints for i in range(counts[0]): keypoints = objects[0].keypoints x = keypoints[:, 2] y = keypoints[:, 1] for j in range(num_parts): if x[j] > 0 and y[j] > 0: cv2.circle(frame, (int(x[j]), int(y[j])), 3, (0, 255, 255), thickness=-1, lineTy pe =cv2.FILLED) for j, e in enumerate(part_edges): if x[e].min() > 0 and y[e].min() > 0: cv2.line(frame, (int(x[e[0]]), int(y[e[0]])), (int(x[e[1]]), int(y[e[1]])), (0, 255, 255), 2) # Display the resulting frame cv2.imshow('frame', frame) if cv2.waitKey(1) == ord('q'): break # When everything done, release the capture cap.release() cv2.destroyAllWindows()