https://developer.nvidia.com/downloads/embedded/l4t/r35_release_v3.1/sd_card_b49/jp511-orin-nano-sd-card-image.zip/

- CUDA11.4.19- cuDNN8.6.0- TensorRT8.5.2- OpenCV4.5.4

https://www.balena.io/etcher#download-etcher

烧录完成以后插到Jetson Orin Nano开发板的风扇下方的卡槽中，图示如下：

https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform/index.html

sudo apt-get -y install autoconf bc build-essential g++-8 gcc-8 clang-8 lld-8 gettext-base gfortran-8 iputils-ping libbz2-dev libc++-dev libcgal-dev libffi-dev libfreetype6-dev libhdf5-dev libjpeg-dev liblzma-dev libncurses5-dev libncursesw5-dev libpng-dev libreadline-dev libssl-dev libsqlite3-dev libxml2-dev libxslt-dev locales moreutils openssl python-openssl rsync scons python3-pip libopenblas-dev;

pip3 install torch-1.14.0a0+44dac51c.nv23.02-cp38-cp38-linux_aarch64.whl

sudo apt install libjpeg-dev zlib1g-dev libpython3-dev libavcodec-dev libavformat-dev libswscale-devpip3 install --upgrade pillowwget https://github.com/pytorch/vision/archive/refs/tags/v0.15.1.zipunzip v0.15.1.zipcd vision-0.15.1export BUILD_VERSION=0.15.1python3 setup.py install --user

说明一切准备工作就绪了。

说明一下，安装过程中要求输入提示的都输入 y

cd /usr/src/tensorrt/bin./trtexec --onnx=<path_to_onnx_file> --saveEngine=<path_to_save_engine_file>

这个时间大概在五分钟左右，需要等一下才可以转换好。

  1import tensorrt as trt
  2from torchvision import transforms
  3import torch as t
  4from collections import OrderedDict, namedtuple
  5import cv2 as cv
  6import time
  7import numpy as np
  8
  9img_transform = transforms.Compose([transforms.ToTensor(),
 10                                    transforms.Resize((640, 640))
 11                                    ])
 12
 13def load_classes():
 14    with open("classes.txt", "r") as f:
 15        class_list = [cname.strip() for cname in f.readlines()]
 16    return class_list
 17
 18
 19def format_yolov8(frame):
 20    row, col, _ = frame.shape
 21    _max = max(col, row)
 22    result = np.zeros((_max, _max, 3), np.uint8)
 23    result[0:row, 0:col] = frame
 24    result = cv.cvtColor(result, cv.COLOR_BGR2RGB)
 25    return result
 26
 27def wrap_detection(input_image, output_data):
 28    class_ids = []
 29    confidences = []
 30    boxes = []
 31    out_data = output_data.T
 32    rows = out_data.shape[0]
 33
 34    image_width, image_height, _ = input_image.shape
 35
 36    x_factor = image_width / 640.0
 37    y_factor = image_height / 640.0
 38
 39    for r in range(rows):
 40        row = out_data[r]
 41        classes_scores = row[4:]
 42        class_id = np.argmax(classes_scores)
 43        if (classes_scores[class_id] > .25):
 44            class_ids.append(class_id)
 45            confidences.append(classes_scores[class_id])
 46            x, y, w, h = row[0].item(), row[1].item(), row[2].item(), row[3].item()
 47            left = int((x - 0.5 * w) * x_factor)
 48            top = int((y - 0.5 * h) * y_factor)
 49            width = int(w * x_factor)
 50            height = int(h * y_factor)
 51            box = np.array([left, top, width, height])
 52            boxes.append(box)
 53
 54    indexes = cv.dnn.NMSBoxes(boxes, confidences, 0.25, 0.25)
 55
 56    result_class_ids = []
 57    result_confidences = []
 58    result_boxes = []
 59
 60    for i in indexes:
 61        result_confidences.append(confidences[i])
 62        result_class_ids.append(class_ids[i])
 63        result_boxes.append(boxes[i])
 64
 65    return result_class_ids, result_confidences, result_boxes
 66def gpu_trt_demo():
 67    class_list = load_classes()
 68    device = t.device('cuda:0')
 69    Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
 70    logger = trt.Logger(trt.Logger.INFO)
 71    with open("yolov8n.engine", 'rb') as f, trt.Runtime(logger) as runtime:
 72        model = runtime.deserialize_cuda_engine(f.read())
 73    bindings = OrderedDict()
 74    for index in range(model.num_bindings):
 75        name = model.get_binding_name(index)
 76        dtype = trt.nptype(model.get_binding_dtype(index))
 77        shape = model.get_binding_shape(index)
 78        data = t.from_numpy(np.empty(shape, dtype=np.dtype(dtype))).to(device)
 79        bindings[name] = Binding(name, dtype, shape, data, int(data.data_ptr()))
 80    binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items())
 81    context = model.create_execution_context()
 82
 83    capture = cv.VideoCapture("test.mp4")
 84    colors = [(255, 255, 0), (0, 255, 0), (0, 255, 255), (255, 0, 0)]
 85    while True:
 86        _, frame = capture.read()
 87        if frame is None:
 88            print("End of stream")
 89            break
 90        fh, fw, fc = frame.shape
 91        start = time.time()
 92        image = format_yolov8(frame)
 93        x_input = img_transform(image).view(1, 3, 640, 640).to(device)
 94        binding_addrs['images'] = int(x_input.data_ptr())
 95        context.execute_v2(list(binding_addrs.values()))
 96        out_prob = bindings['output0'].data.cpu().numpy()
 97        end = time.time()
 98
 99        class_ids, confidences, boxes = wrap_detection(image, np.squeeze(out_prob, 0))
100        for (classid, confidence, box) in zip(class_ids, confidences, boxes):
101            if box[2] > fw * 0.67:
102                continue
103            color = colors[int(classid) % len(colors)]
104            cv.rectangle(frame, box, color, 2)
105            cv.rectangle(frame, (box[0], box[1] - 20), (box[0] + box[2], box[1]), color, -1)
106            cv.putText(frame, class_list[classid] + " " + ("%.2f"%confidence), (box[0], box[1] - 10), cv.FONT_HERSHEY_SIMPLEX, .5, (0, 0, 0))
107
108        inf_end = end - start
109        fps = 1 / inf_end
110        fps_label = "FPS: %.2f" % fps
111        cv.putText(frame, fps_label, (10, 25), cv.FONT_HERSHEY_SIMPLEX, 1, (0, 0, 255), 2)
112        cv.imshow("YOLOv8 + TensorRT8.5.x Object Detection", frame)
113        cc = cv.waitKey(1)
114        if cc == 27:
115            break
116    cv.waitKey(0)
117    cv.destroyAllWindows()
118
119
120if __name__ == "__main__":
121    gpu_trt_demo()

总结：