奥比中光基于yolov算法进行目标关键特征点识别及深度点云提取

研究了一阵子的yolov5算法，个人是利用CPU跑算法，训练速度很慢，效果也一般，需要自己训练yolov算法，并将算法封装，提供API接口，供其调用。环境配置和代码封装见文末。
import math
import cv2
import detect_with_API
import torch
from openni import openni2
import numpy as np
def depth_img(depth_dpt, x1, y1, x3, y3):
    PointCloudData1 = []
    PointCloudData3 = []
    # 相机内参
    fx = 579.446
    fy = 579.1467
    cx = 318.1348
    cy = 239.8945
    # 对每个点的像素深度值进行处理，获取的是平面数据，再根据平面坐标xy确定z的数据，Z值有误差，需要用陀螺仪校准，特征会漂移
    # print('x1= ', x1)
    # print('y1= ', y1)
    for x in range(0, len(depth_dpt), 1):  # np数组中获取高度值，  遍历每一列 每个高度里面有多少个像素点
        for y in range(0, len(depth_dpt[x])):  # 遍历每一行
            if (depth_dpt[x][y] != 0):  # 判定每个像素点的深度值
                if x in y1 and y in x1:
                    for i in range(len(x1)):
                        if x == y1[i] and y == x1[i]:
                            # PointCloudData_2.append([y, x, depth_dpt[x, y]])  # 存储 y 、 x 、z坐标
                            x_point = float((y - cx) * (depth_dpt[x, y])) / fx
                            y_point = float((x - cy) * (depth_dpt[x, y])) / fy
                            # PointCloudData.append(f'{x_point} {y_point} {depth_dpt[x][y]}\n')
                            PointCloudData1.append([x_point, y_point, depth_dpt[x, y]])  # 存储 y 、 x 、z坐标
                elif x in y3 and y in x3:
                    for i in range(len(x3)):
                        if x == y3[i] and y == x3[i]:
                            # PointCloudData_2.append([y, x, depth_dpt[x, y]])  # 存储 y 、 x 、z坐标
                            x_point = float((y - cx) * (depth_dpt[x, y])) / fx
                            y_point = float((x - cy) * (depth_dpt[x, y])) / fy
                            # PointCloudData.append(f'{x_point} {y_point} {depth_dpt[x][y]}\n')
                            PointCloudData3.append([x_point, y_point, depth_dpt[x, y]])  # 存储 y 、 x 、z坐标
                            cv2.calibrateHandEye()
                            cv2.solvePnP()
    print('特征点point1真实三维坐标：', PointCloudData1)
    print('特征点point3真实三维坐标：', PointCloudData3)
    print()  # 将每一帧的结果输出分开
    angle = (PointCloudData1[0][0] - PointCloudData3[0][0]) / math.sqrt(math.pow((PointCloudData1[0][0] - PointCloudData3[0][0]),2)+math.pow((PointCloudData1[0][2] - PointCloudData3[0][2]),2))
    # print('y坐标：', PointCloudData_2[1])
    # print('z坐标：', PointCloudData_2[2])
    # with open('data.txt', 'w') as F:
    #     F.writelines(PointCloudData)
    return angle
if __name__ == '__main__':
    cap = cv2.VideoCapture(0)
    a = detect_with_API.DetectAPI(weights='runs/train/exp5/weights/best.pt')
    with torch.no_grad():
        rec, img = cap.read()
        result, names = a.detect([img])
        img = result[0][0]  # 每一帧图片的处理结果图片
        # 每一帧图像的识别结果（可包含多个物体）
        x_point1 = []
        y_point1 = []
        x_point2 = []
        y_point2 = []
        x_point3 = []
        y_point3 = []
        for cls, (x1, y1, x2, y2), conf in result[0][1]:
            # print(names[cls], x1, y1, x2, y2, conf)  # 识别物体种类、左上角x坐标、左上角y轴坐标、右下角x轴坐标、右下角y轴坐标，置信度
            x0 = int(x1 + (x2 - x1) / 2)
            y0 = int(y1 + (y2 - y1) / 2)
            if names[cls] == 'point1':
                x_point1.append(x0)
                y_point1.append(y0)
            if names[cls] == 'point3':
                x_point2.append(x0)
                y_point2.append(y0)
            '''
            # cv2.rectangle(img,(x1,y1),(x2,y2),(0,255,0))
            # cv2.putText(img,names[cls],(x1,y1-20),cv2.FONT_HERSHEY_DUPLEX,1.5,(255,0,0))'''
        cap.release()
        cv2.destroyAllWindows()
        openni2.initialize()
        dev = openni2.Device.open_any()
        print(dev.get_device_info())
        depth_stream = dev.create_depth_stream()
        depth_stream.start()
        frame = depth_stream.read_frame()
        dframe_data = np.array(frame.get_buffer_as_triplet()).reshape([480, 640, 2])
        dpt1 = np.asarray(dframe_data[:, :, 0], dtype='float32')
        dpt2 = np.asarray(dframe_data[:, :, 1], dtype='float32')
        dpt2 *= 255
        dpt = dpt1 + dpt2
        dpt = dpt[:, ::-1]
        angle = depth_img(dpt, x_point1, y_point1, x_point3, y_point3)
        depth_stream.stop()
        dev.close()

yolov5环境配置和代码封装都有很详细的教程，下面提供参考链接：

yolov5环境配置：在CPU上跑yolov5（详细步骤+适合入门）_yolo 怎么跑起来-CSDN博客
【Yolov5】1.认真总结6000字Yolov5保姆级教程（2022.06.28全新版本v6.1）_yolov5教程-CSDN博客

yolov5代码封装：魔改并封装 YoloV5 Version7 的 detect.py 成 API接口以供 python 程序使用_yolov5部署成api接口-CSDN博客