YOLOv7教程系列：三、使用onnx导出成openvino模型，并调用接口进行摄像头推理预测_onnx转openvino

0.引言

OpenVINO是英特尔推出的一款全面的工具套件，用于快速部署应用和解决方案，支持计算机视觉的CNN网络结构超过200余种。openvino官方文档
本文将介绍使用yolov7+openvino在cpu上快速部署检测模型，如果还不会训练的朋友可以移步至另一篇博客：
YOLOv7系列教程：一、基于自定义数据集训练专属于自己的目标检测模型（保姆级教程，含数据集预处理），包含对train.py/test.py/detect.py/export.py详细说明
项目在github上开源，项目地址：yolov7_openvino_python

1.安装环境

进入自己的虚拟环境，使用pip快速安装，安装成功即可进入下一步：

pip install -r requirements.txt
1

安装完成截图：

2.转换模型

使用mo工具，可以快速将onnx中间模型导出成openvino模型，目前只试过p5的模型进行openvino部署，具体可以参考mo使用说明

mo --input_model weights/best.onnx --output_dir yolov7_openvino
1

–input_model是输入的onnx模型，路径改成自己的路径
–output_dir是输出文件夹，此时会生成一个yolov7_openvino文件夹 ，并将.xml/.bin/.mapping文件保存于此

转换完成截图：

3.运行代码

先确保主机上有板载摄像头或者外接USB摄像头，运行如下命令即可：

python detect.py
1

运行成功截图：

4.代码详情

（1）主代码

主要是配置变量，并且进行opencv读图循环

import os
import cv2

from model.yolov7 import YOLOV7_OPENVINO



if __name__ == "__main__":
    class_list = [
        "person", "bicycle", "car", "motorcycle", "airplane", "bus", "train", "truck", "boat", "traffic light",
        "fire hydrant", "stop sign", "parking meter", "bench", "bird", "cat", "dog", "horse", "sheep", "cow",
        "elephant", "bear", "zebra", "giraffe", "backpack", "umbrella", "handbag", "tie", "suitcase", "frisbee",
        "skis", "snowboard", "sports ball", "kite", "baseball bat", "baseball glove", "skateboard", "surfboard",
        "tennis racket", "bottle", "wine glass", "cup", "fork", "knife", "spoon", "bowl", "banana", "apple",
        "sandwich", "orange", "broccoli", "carrot", "hot dog", "pizza", "donut", "cake", "chair", "couch",
        "potted plant", "bed", "dining table", "toilet", "tv", "laptop", "mouse", "remote", "keyboard", "cell phone",
        "microwave", "oven", "toaster", "sink", "refrigerator", "book", "clock", "vase", "scissors", "teddy bear",
        "hair drier", "toothbrush"
       ]
    ov_name = "yolov7_openvino/best.xml"
    ov_path = os.path.join(os.getcwd(), "model", ov_name)
    yolov7_detector=YOLOV7_OPENVINO(class_list, ov_path)


    cap = cv2.VideoCapture(0)
    while True:
        # get a frame
        ret, frame = cap.read()
        # infer
        yolov7_detector.infer_image(frame)
        # show a frame
        cv2.imshow("capture", frame)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break
    cap.release()
    cv2.destroyAllWindows()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36

（2）yolov7_openvino版代码

from openvino.runtime import Core
import cv2
import numpy as np
import random
import time
from openvino.preprocess import PrePostProcessor, ColorFormat
from openvino.runtime import Layout, AsyncInferQueue, PartialShape

class YOLOV7_OPENVINO(object):
    def __init__(self, class_list, model_path):
        # set the hyperparameters
        self.classes = class_list
        self.batchsize = 1
        self.grid = False
        self.img_size = (640, 640) 
        self.conf_thres = 0.5
        self.iou_thres = 0.6
        self.class_num = len(self.classes)
        self.colors = [[random.randint(0, 255) for _ in range(3)] for _ in self.classes]
        self.stride = [8, 16, 32]
        self.anchor_list = [[12, 16, 19, 36, 40, 28], [36, 75, 76, 55, 72, 146], [142, 110, 192, 243, 459, 401]]
        self.anchor = np.array(self.anchor_list).astype(float).reshape(3, -1, 2)
        area = self.img_size[0] * self.img_size[1]
        device = 'CPU'
        nireq = 2
        self.size = [int(area / self.stride[0] ** 2), int(area / self.stride[1] ** 2), int(area / self.stride[2] ** 2)]
        self.feature = [[int(j / self.stride[i]) for j in self.img_size] for i in range(3)]

        ie = Core()
        self.model = ie.read_model(model_path)
        self.input_layer = self.model.input(0)
        new_shape = PartialShape([self.batchsize, 3, self.img_size[0], self.img_size[1]])
        self.model.reshape({self.input_layer.any_name: new_shape})
        self.pre_api = True
        if (self.pre_api == True):
            # Preprocessing API
            ppp = PrePostProcessor(self.model)
            # Declare section of desired application's input format
            ppp.input().tensor() \
                .set_layout(Layout("NHWC")) \
                .set_color_format(ColorFormat.BGR)
            # Here, it is assumed that the model has "NCHW" layout for input.
            ppp.input().model().set_layout(Layout("NCHW"))
            # Convert current color format (BGR) to RGB
            ppp.input().preprocess() \
                .convert_color(ColorFormat.RGB) \
                .scale([255.0, 255.0, 255.0])
            self.model = ppp.build()
            print(f'Dump preprocessor: {ppp}')

        self.compiled_model = ie.compile_model(model=self.model, device_name=device)
        self.infer_queue = AsyncInferQueue(self.compiled_model, nireq)

    def letterbox(self, img, new_shape=(640, 640), color=(114, 114, 114)):
        # Resize and pad image while meeting stride-multiple constraints
        shape = img.shape[:2]  # current shape [height, width]
        if isinstance(new_shape, int):
            new_shape = (new_shape, new_shape)

        # Scale ratio (new / old)
        r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
        new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
        dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - \
            new_unpad[1]  # wh padding

        # divide padding into 2 sides
        dw /= 2
        dh /= 2

        # resize
        if shape[::-1] != new_unpad:
            img = cv2.resize(img, new_unpad, interpolation=cv2.INTER_LINEAR)
        top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
        left, right = int(round(dw - 0.1)), int(round(dw + 0.1))

        # add border
        img = cv2.copyMakeBorder(img, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color)

        return img

    def xywh2xyxy(self, x):
        # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
        y = np.copy(x)
        y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
        y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
        y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
        y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y

        return y

    def nms(self, prediction, conf_thres, iou_thres):
        predictions = np.squeeze(prediction[0])

        # Filter out object confidence scores below threshold
        obj_conf = predictions[:, 4]
        predictions = predictions[obj_conf > conf_thres]
        obj_conf = obj_conf[obj_conf > conf_thres]

        # Multiply class confidence with bounding box confidence
        predictions[:, 5:] *= obj_conf[:, np.newaxis]

        # Get the scores
        scores = np.max(predictions[:, 5:], axis=1)

        # Filter out the objects with a low score
        valid_scores = scores > conf_thres
        predictions = predictions[valid_scores]
        scores = scores[valid_scores]

        # Get the class with the highest confidence
        class_ids = np.argmax(predictions[:, 5:], axis=1)

        # Get bounding boxes for each object
        boxes = self.xywh2xyxy(predictions[:, :4])

        # Apply non-maxima suppression to suppress weak, overlapping bounding boxes
        # indices = nms(boxes, scores, self.iou_threshold)
        indices = cv2.dnn.NMSBoxes(boxes.tolist(), scores.tolist(), conf_thres, iou_thres)

        return boxes[indices], scores[indices], class_ids[indices]

    def clip_coords(self, boxes, img_shape):
        # Clip bounding xyxy bounding boxes to image shape (height, width)
        boxes[:, 0].clip(0, img_shape[1])  # x1
        boxes[:, 1].clip(0, img_shape[0])  # y1
        boxes[:, 2].clip(0, img_shape[1])  # x2
        boxes[:, 3].clip(0, img_shape[0])  # y2

    def scale_coords(self, img1_shape, img0_shape, coords, ratio_pad=None):
        # Rescale coords (xyxy) from img1_shape to img0_shape
        # gain  = old / new
        if ratio_pad is None:
            gain = min(img1_shape[0] / img0_shape[0],
                       img1_shape[1] / img0_shape[1])
            padding = (img1_shape[1] - img0_shape[1] * gain) / \
                2, (img1_shape[0] - img0_shape[0] * gain) / 2
        else:
            gain = ratio_pad[0][0]
            padding = ratio_pad[1]
        coords[:, [0, 2]] -= padding[0]  # x padding
        coords[:, [1, 3]] -= padding[1]  # y padding
        coords[:, :4] /= gain
        self.clip_coords(coords, img0_shape)

    def sigmoid(self, x):
        return 1 / (1 + np.exp(-x))

    def plot_one_box(self, x, img, color=None, label=None, line_thickness=None):
        # Plots one bounding box on image img
        tl = line_thickness or round(
            0.002 * (img.shape[0] + img.shape[1]) / 2) + 1  # line/font thickness
        color = color or [random.randint(0, 255) for _ in range(3)]
        c1, c2 = (int(x[0]), int(x[1])), (int(x[2]), int(x[3]))
        cv2.rectangle(img, c1, c2, color, thickness=tl, lineType=cv2.LINE_AA)
        if label:
            tf = max(tl - 1, 1)  # font thickness
            t_size = cv2.getTextSize(
                label, 0, fontScale=tl / 3, thickness=tf)[0]
            c2 = c1[0] + t_size[0], c1[1] - t_size[1] - 3
            cv2.rectangle(img, c1, c2, color, -1, cv2.LINE_AA)  # filled
            cv2.putText(img, label, (c1[0], c1[1] - 2), 0, tl / 3,
                        [225, 255, 255], thickness=tf, lineType=cv2.LINE_AA)
    
    def draw(self, img, boxinfo):
        for i, (xyxy, conf, cls) in enumerate(boxinfo):
            self.plot_one_box(xyxy, img, label=self.classes[int(cls)], color=self.colors[int(cls)], line_thickness=2)
            # cv2.putText()
            print(i)
        # cv2.imshow('Press ESC to Exit', img)
        # cv2.waitKey(1)

    def postprocess(self, infer_request, info):
        t3 = time.time()
        src_img_list, src_size = info
        for batch_id in range(self.batchsize):
            if self.grid:
                results = np.expand_dims(infer_request.get_output_tensor(0).data[batch_id], axis=0)
            else:
                output = []
                # Get the each feature map's output data
                output.append(self.sigmoid(infer_request.get_output_tensor(0).data[batch_id].reshape(-1, self.size[0]*3, 5+self.class_num)))
                output.append(self.sigmoid(infer_request.get_output_tensor(1).data[batch_id].reshape(-1, self.size[1]*3, 5+self.class_num)))
                output.append(self.sigmoid(infer_request.get_output_tensor(2).data[batch_id].reshape(-1, self.size[2]*3, 5+self.class_num)))
                
                # Postprocessing
                grid = []
                for _, f in enumerate(self.feature):
                    grid.append([[i, j] for j in range(f[0]) for i in range(f[1])])

                result = []
                for i in range(3):
                    src = output[i]
                    xy = src[..., 0:2] * 2. - 0.5
                    wh = (src[..., 2:4] * 2) ** 2
                    dst_xy = []
                    dst_wh = []
                    for j in range(3):
                        dst_xy.append((xy[:, j * self.size[i]:(j + 1) * self.size[i], :] + grid[i]) * self.stride[i])
                        dst_wh.append(wh[:, j * self.size[i]:(j + 1) *self.size[i], :] * self.anchor[i][j])
                    src[..., 0:2] = np.concatenate((dst_xy[0], dst_xy[1], dst_xy[2]), axis=1)
                    src[..., 2:4] = np.concatenate((dst_wh[0], dst_wh[1], dst_wh[2]), axis=1)
                    result.append(src)
                results = np.concatenate(result, 1)
                
            boxes, scores, class_ids = self.nms(results, self.conf_thres, self.iou_thres)
            img_shape = self.img_size
            self.scale_coords(img_shape, src_size, boxes)

            # Draw the results
            self.draw(src_img_list[batch_id], zip(boxes, scores, class_ids))
        t4 = time.time()
        print(f"post time{(t4-t3)*1000}")

    def infer_image(self, src_img):

        src_img_list = []
        src_img_list.append(src_img)
        img = self.letterbox(src_img, self.img_size)
        src_size = src_img.shape[:2]
        img = img.astype(dtype=np.float32)
        if (self.pre_api == False):
            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)  # BGR to RGB
            img /= 255.0
            img.transpose(2, 0, 1) # NHWC to NCHW
        input_image = np.expand_dims(img, 0)

        # Set callback function for postprocess
        self.infer_queue.set_callback(self.postprocess)
        # Do inference
        self.infer_queue.start_async({self.input_layer.any_name: input_image}, (src_img_list, src_size))

        self.infer_queue.wait_all()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232