opencv 怎么找角度 opencv角度识别

模型介绍

OpenVINO支持头部姿态评估模型，预训练模型为：head-pose-estimation-adas-0001，在三个维度方向实现头部动作识别，它们分别是：

pitch是俯仰角，是“点头“
yaw是偏航角，是‘摇头’
roll是旋转角，是“翻滚

它们的角度范围分别为：

YAW [-90,90], PITCH [-70,70], ROLL [-70,70]

这三个专业词汇其实是来自无人机与航空领域，计算机视觉科学家一大爱好就是搞新词，就把它们借用到头部姿态评估中，它们的意思图示如下：

对应到头部姿态评估中

输入格式：[1x3x60x60] BGR顺序

输出格式：

name: "angle_y_fc", shape: [1, 1] - Estimated
name: "angle_p_fc", shape: [1, 1] - Estimated pitch
name: "angle_r_fc", shape: [1, 1] - Estimated roll

人脸检测

基于OpenVINO中MobileNetv2 SSD人脸检测模型，实现人脸检测，然后得到ROI区域，基于ROI实现头部姿态评估，完成头部动作识别，这里只会识别幅度超过正负20度以上的头部动作。实现模型加载与输入输出格式解析的代码如下：

1ie = IECore()
 2for device in ie.available_devices:
 3    print(device)
 4
 5net = ie.read_network(model=model_xml, weights=model_bin)
 6input_blob = next(iter(net.input_info))
 7out_blob = next(iter(net.outputs))
 8
 9n, c, h, w = net.input_info[input_blob].input_data.shape
10print(n, c, h, w)
11
12# cap = cv.VideoCapture("D:/images/video/Boogie_Up.mp4")
13cap = cv.VideoCapture("D:/images/video/example_dsh.mp4")
14# cap = cv.VideoCapture(0)
15exec_net = ie.load_network(network=net, device_name="CPU")
16
17em_net = ie.read_network(model=em_xml, weights=em_bin)
18em_input_blob = next(iter(em_net.input_info))
19em_it = iter(em_net.outputs)
20em_out_blob1 = next(em_it)  # angle_y_fc
21em_out_blob2 = next(em_it)  # angle_p_fc
22em_out_blob3 = next(em_it)  # angle_r_fc
23print(em_out_blob1, em_out_blob2, em_out_blob3)
24en, ec, eh, ew = em_net.input_info[em_input_blob].input_data.shape
25print(en, ec, eh, ew)
26
27em_exec_net = ie.load_network(network=em_net, device_name="CPU"

实现头部动作检测

解析模型的输出，对视频流实现人脸检测与头部动作识别的代码如下：

1height = cap.get(cv.CAP_PROP_FRAME_HEIGHT)
 2width = cap.get(cv.CAP_PROP_FRAME_WIDTH)
 3count = cap.get(cv.CAP_PROP_FRAME_COUNT)
 4fps = cap.get(cv.CAP_PROP_FPS)
 5out = cv.VideoWriter("D:/test.mp4", cv.VideoWriter_fourcc('D', 'I', 'V', 'X'), 15, (np.int(width), np.int(height)),
 6                     True)
 7while True:
 8    ret, frame = cap.read()
 9    if ret is not True:
10        break
11    image = cv.resize(frame, (w, h))
12    image = image.transpose(2, 0, 1)
13    inf_start = time.time()
14    res = exec_net.infer(inputs={input_blob: [image]})
15    inf_end = time.time() - inf_start
16    # print("infer time(ms)：%.3f"%(inf_end*1000))
17    ih, iw, ic = frame.shape
18    res = res[out_blob]
19    for obj in res[0][0]:
20        if obj[2] > 0.75:
21            xmin = int(obj[3] * iw)-10
22            ymin = int(obj[4] * ih)-10
23            xmax = int(obj[5] * iw)+10
24            ymax = int(obj[6] * ih)+10
25            if xmin < 0:
26                xmin = 0
27            if ymin < 0:
28                ymin = 0
29            if xmax >= iw:
30                xmax = iw - 1
31            if ymax >= ih:
32                ymax = ih - 1
33            roi = frame[ymin:ymax, xmin:xmax, :]
34            roi_img = cv.resize(roi, (ew, eh))
35            roi_img = roi_img.transpose(2, 0, 1)
36            em_res = em_exec_net.infer(inputs={em_input_blob: [roi_img]})
37            angle_p_fc = em_res[em_out_blob1][0][0]
38            angle_r_fc = em_res[em_out_blob2][0][0]
39            angle_y_fc = em_res[em_out_blob3][0][0]
40            postxt = ""
41            if angle_p_fc > 10 or angle_p_fc < -10:
42                postxt += "pitch, "
43            if angle_y_fc > 10 or angle_y_fc < -10:
44                postxt += "yaw, "
45            if angle_r_fc > 10 or angle_r_fc < -10:
46                postxt += "roll, "
47
48            cv.putText(frame, postxt, (xmin, ymin-10), cv.FONT_HERSHEY_SIMPLEX, 1.0, (0, 0, 255), 2)
49            cv.rectangle(frame, (xmin, ymin), (xmax, ymax), (0, 255, 255), 2, 8)
50            cv.putText(frame, "infer time(ms): %.3f" % (inf_end * 1000), (50, 50), cv.FONT_HERSHEY_SIMPLEX, 1.0,
51                       (255, 0, 255),
52                       2, 8)
53    cv.imshow("Face & head pose demo", frame)
54    out.write(frame)
55    c = cv.waitKey(1)
56    if c == 27:
57        break
58cv.waitKey(0)
59out.release()
60cap.release()

运行结果如下：

视频文件

这个建议感兴趣的可以尝试一下，把视频文件换成摄像头，基本上实时识别点头，摇头，转头等动作毫无压力，我自所以不用我自己的测试截图，主要把我自己长的太丑！另外速度实时！真的好用！