Windows 11 配置 Intel RealSense D435i 深度相机及 Python 开发

数据采集与保存

在实际项目中，往往需要离线存储图像数据。下面是一个简单的 RGB 采集脚本，支持按快捷键保存。

import os
import cv2
import numpy as np
import pyrealsense2 as rs
import time

# 初始化 Pipeline
pipeline = rs.pipeline()
config = rs.config()
config.enable_stream(rs.stream.color, 1920, 1080, rs.format.bgr8, 15)
profile = pipeline.start(config)

cv2.namedWindow('RGB Image', cv2.WINDOW_AUTOSIZE)
save_dir = 'E:/my_images/'
os.makedirs(save_dir, exist_ok=True)
counter = 0

try:
    while True:
        frames = pipeline.wait_for_frames()
        color_frame = frames.get_color_frame()
        if not color_frame:
            continue
        
        color_image = np.asanyarray(color_frame.get_data())
        cv2.imshow('RGB Image', color_image)
        
        key = cv2.waitKey(1)
        if key == ord('t'):  # 按 T 键保存
            rgb_file = os.path.join(save_dir, f'rgb_{counter:04d}.png')
            cv2.imwrite(rgb_file, color_image)
            print(f'Saved: {rgb_file}')
            counter += 1
        elif key in [27, ord('q')]:  # ESC 或 Q 退出
            break
finally:
    pipeline.stop()

进阶应用：YOLOv8 与关键点识别

1. YOLOv8 目标检测与三维坐标

将深度信息映射到检测框内，可获取目标的精确距离和空间位置。关键在于使用 rs.align 将深度流对齐到彩色流，解决视差问题。

import cv2
import pyrealsense2 as rs
import numpy as np
from ultralytics import YOLO

model = YOLO("runs/detect/train/weights/best.pt")

pipeline = rs.pipeline()
config = rs.config()
config.enable_stream(rs.stream.depth, 848, 480, rs.format.z16, 30)
config.enable_stream(rs.stream.color, 848, 480, rs.format.bgr8, 30)
pipeline.start(config)

align_to = rs.stream.color
align = rs.align(align_to)

def get_aligned_images():
    frames = pipeline.wait_for_frames()
    aligned_frames = align.process(frames)
    
    aligned_depth_frame = aligned_frames.get_depth_frame()
    color_frame = aligned_frames.get_color_frame()
    
    intr = color_frame.profile.as_video_stream_profile().intrinsics
    depth_intrin = aligned_depth_frame.profile.as_video_stream_profile().intrinsics
    
    depth_image = np.asanyarray(aligned_depth_frame.get_data())
    depth_image_8bit = cv2.convertScaleAbs(depth_image, alpha=0.03)
    depth_image_3d = np.dstack((depth_image_8bit, depth_image_8bit, depth_image_8bit))
    color_image = np.asanyarray(color_frame.get_data())
    
    return intr, depth_intrin, color_image, depth_image, aligned_depth_frame

def get_3d_camera_coordinate(depth_pixel, aligned_depth_frame, depth_intrin):
    x, y = depth_pixel
    dis = aligned_depth_frame.get_distance(x, y)
    camera_coordinate = rs.rs2_deproject_pixel_to_point(depth_intrin, depth_pixel, dis)
    return dis, camera_coordinate

try:
    while True:
        intr, depth_intrin, color_image, depth_image, aligned_depth_frame = get_aligned_images()
        if not depth_image.any() or not color_image.any():
            continue
        
        results = model.predict(color_image, conf=0.5)
        annotated_frame = results[0].plot()
        detected_boxes = results[0].boxes.xyxy
        
        for i, box in enumerate(detected_boxes):
            x1, y1, x2, y2 = map(int, box)
            xrange = max(1, int(abs((x1 - x2) / 30)))
            yrange = max(1, int(abs((y1 - y2) / 30)))
            
            point_cloud_data = []
            for xp in range(x1, x2, xrange):
                for yp in range(y1, y2, yrange):
                    dis, coord = get_3d_camera_coordinate([xp, yp], aligned_depth_frame, depth_intrin)
                    point_cloud_data.append(f"{coord}")
            
            with open("point_cloud_data.txt", "a") as file:
                file.write(f"\nTime: {time.time()}\n")
                file.write(" ".join(point_cloud_data))
            
            ux, uy = int((x1 + x2) / 2), int((y1 + y2) / 2)
            dis, coord = get_3d_camera_coordinate([ux, uy], aligned_depth_frame, depth_intrin)
            formatted_coord = f"({coord[0]:.2f}, {coord[1]:.2f}, {coord[2]:.2f})"
            
            cv2.circle(annotated_frame, (ux, uy), 4, (255, 255, 255), 5)
            cv2.putText(annotated_frame, formatted_coord, (ux + 20, uy + 10), 
                        0, 1, [225, 255, 255], thickness=1, lineType=cv2.LINE_AA)
        
        cv2.imshow('YOLOv8 RealSense', annotated_frame)
        if cv2.waitKey(1) & 0xFF == ord('q'):
            break
finally:
    pipeline.stop()

2. 关键点识别逻辑

对于姿态估计任务，流程类似：先获取对齐后的深度与彩色帧，提取相机内参，再遍历关键点坐标计算深度值。

1. 获取对齐帧：使用 align.process(frames) 确保深度与颜色像素一一对应。
2. 提取内参：通过 profile.as_video_stream_profile().intrinsics 获取焦距和主点。
3. 坐标转换：利用 get_distance 获取像素深度，结合内参反投影至相机坐标系。

这种方案能有效融合 2D 视觉感知与 3D 空间信息，适用于机器人抓取、人体动作分析等场景。