yolov8直接调用zed相机实现三维测距(python)

yolov8直接调用zed相机实现三维测距(python)

  • [1. 相关配置](#1. 相关配置)
  • [2. 版本一](#2. 版本一)
    • [2.1 相关代码](#2.1 相关代码)
    • [2.2 实验结果](#2.2 实验结果)
  • [3. 版本二](#3. 版本二)
    • [3.1 相关代码](#3.1 相关代码)
    • [3.2 实验结果](#3.2 实验结果)

相关链接

此项目直接调用zed相机实现三维测距,无需标定,相关内容如下:
1.yolov5直接调用zed相机实现三维测距(python)
2. yolov4直接调用zed相机实现三维测距
3. Windows+YOLOV8环境配置
4.具体实现效果已在哔哩哔哩发布,点击此链接跳转

本篇博文工程源码下载(麻烦github给个星星)
下载链接:https://github.com/up-up-up-up/zed-yolov8

附:Zed调用YOLOv7测距也已经实现,但是3060笔记本6G显存带不动,在大现存服务器上可以运行,可能是由于YOLOv7网络结构导致的,由于不具备普适性,就不再写相关文章了,有需要的可以仿照这个代码去改写

1. 相关配置

python==3.7

Windows-pycharm

zed api 具体配置见 (zed api 配置步骤)

由于我电脑之前python版本为3.7,yolov8要求python最低为3.8,所以本次实验直接在虚拟环境里进行,未配置gpu,可能看着卡卡的,有需要的可以配置一下,原理是一样的

2. 版本一

2.1 相关代码

主代码 zed-yolo.py,具体放置在yolov8主目录下,盒子形式展现,可实现测距+跟踪

python 复制代码
#!/usr/bin/env python3

import sys
import numpy as np

import argparse
import torch
import cv2
import pyzed.sl as sl
from ultralytics import YOLO

from threading import Lock, Thread
from time import sleep

import ogl_viewer.viewer as gl
import cv_viewer.tracking_viewer as cv_viewer

lock = Lock()
run_signal = False
exit_signal = False


def xywh2abcd(xywh, im_shape):
    output = np.zeros((4, 2))

    # Center / Width / Height -> BBox corners coordinates
    x_min = (xywh[0] - 0.5*xywh[2]) #* im_shape[1]
    x_max = (xywh[0] + 0.5*xywh[2]) #* im_shape[1]
    y_min = (xywh[1] - 0.5*xywh[3]) #* im_shape[0]
    y_max = (xywh[1] + 0.5*xywh[3]) #* im_shape[0]

    # A ------ B
    # | Object |
    # D ------ C

    output[0][0] = x_min
    output[0][1] = y_min

    output[1][0] = x_max
    output[1][1] = y_min

    output[2][0] = x_max
    output[2][1] = y_max

    output[3][0] = x_min
    output[3][1] = y_max
    return output

def detections_to_custom_box(detections, im0):
    output = []
    for i, det in enumerate(detections):
        xywh = det.xywh[0]

        # Creating ingestable objects for the ZED SDK
        obj = sl.CustomBoxObjectData()
        obj.bounding_box_2d = xywh2abcd(xywh, im0.shape)
        obj.label = det.cls
        obj.probability = det.conf
        obj.is_grounded = False
        output.append(obj)
    return output


def torch_thread(weights, img_size, conf_thres=0.2, iou_thres=0.45):
    global image_net, exit_signal, run_signal, detections

    print("Intializing Network...")

    model = YOLO(weights)

    while not exit_signal:
        if run_signal:
            lock.acquire()

            img = cv2.cvtColor(image_net, cv2.COLOR_BGRA2RGB)
            # https://docs.ultralytics.com/modes/predict/#video-suffixes
            det = model.predict(img, save=False, imgsz=img_size, conf=conf_thres, iou=iou_thres)[0].cpu().numpy().boxes

            # ZED CustomBox format (with inverse letterboxing tf applied)
            detections = detections_to_custom_box(det, image_net)
            lock.release()
            run_signal = False
        sleep(0.01)


def main():
    global image_net, exit_signal, run_signal, detections

    capture_thread = Thread(target=torch_thread, kwargs={'weights': opt.weights, 'img_size': opt.img_size, "conf_thres": opt.conf_thres})
    capture_thread.start()

    print("Initializing Camera...")

    zed = sl.Camera()

    input_type = sl.InputType()
    if opt.svo is not None:
        input_type.set_from_svo_file(opt.svo)

    # Create a InitParameters object and set configuration parameters
    init_params = sl.InitParameters(input_t=input_type, svo_real_time_mode=True)
    init_params.coordinate_units = sl.UNIT.METER
    init_params.depth_mode = sl.DEPTH_MODE.ULTRA  # QUALITY
    init_params.coordinate_system = sl.COORDINATE_SYSTEM.RIGHT_HANDED_Y_UP
    init_params.depth_maximum_distance = 50

    runtime_params = sl.RuntimeParameters()
    status = zed.open(init_params)

    if status != sl.ERROR_CODE.SUCCESS:
        print(repr(status))
        exit()

    image_left_tmp = sl.Mat()

    print("Initialized Camera")

    positional_tracking_parameters = sl.PositionalTrackingParameters()
    # If the camera is static, uncomment the following line to have better performances and boxes sticked to the ground.
    # positional_tracking_parameters.set_as_static = True
    zed.enable_positional_tracking(positional_tracking_parameters)

    obj_param = sl.ObjectDetectionParameters()
#    obj_param.detection_model = sl.OBJECT_DETECTION_MODEL.CUSTOM_BOX_OBJECTS
    obj_param.enable_tracking = True
    zed.enable_object_detection(obj_param)

    objects = sl.Objects()
    obj_runtime_param = sl.ObjectDetectionRuntimeParameters()

    # Display
    camera_infos = zed.get_camera_information()
    camera_res = camera_infos.camera_resolution
    # Create OpenGL viewer
    viewer = gl.GLViewer()
    point_cloud_res = sl.Resolution(min(camera_res.width, 720), min(camera_res.height, 404))
    point_cloud_render = sl.Mat()
    viewer.init(camera_infos.camera_model, point_cloud_res, obj_param.enable_tracking)
    point_cloud = sl.Mat(point_cloud_res.width, point_cloud_res.height, sl.MAT_TYPE.F32_C4, sl.MEM.CPU)
    image_left = sl.Mat()
    # Utilities for 2D display
    display_resolution = sl.Resolution(min(camera_res.width, 1280), min(camera_res.height, 720))
    image_scale = [display_resolution.width / camera_res.width, display_resolution.height / camera_res.height]
    image_left_ocv = np.full((display_resolution.height, display_resolution.width, 4), [245, 239, 239, 255], np.uint8)

    # # Utilities for tracks view
    # camera_config = camera_infos.camera_configuration
    # tracks_resolution = sl.Resolution(400, display_resolution.height)
    # track_view_generator = cv_viewer.TrackingViewer(tracks_resolution, camera_config.fps, init_params.depth_maximum_distance)
    # track_view_generator.set_camera_calibration(camera_config.calibration_parameters)
    # image_track_ocv = np.zeros((tracks_resolution.height, tracks_resolution.width, 4), np.uint8)
    # Camera pose
    cam_w_pose = sl.Pose()

    while viewer.is_available() and not exit_signal:
        if zed.grab(runtime_params) == sl.ERROR_CODE.SUCCESS:
            # -- Get the image
            lock.acquire()
            zed.retrieve_image(image_left_tmp, sl.VIEW.LEFT)
            image_net = image_left_tmp.get_data()
            lock.release()
            run_signal = True

            # -- Detection running on the other thread
            while run_signal:
                sleep(0.001)

            # Wait for detections
            lock.acquire()
            # -- Ingest detections
            zed.ingest_custom_box_objects(detections)
            lock.release()
            zed.retrieve_objects(objects, obj_runtime_param)

            # -- Display
            # Retrieve display data
            zed.retrieve_measure(point_cloud, sl.MEASURE.XYZRGBA, sl.MEM.CPU, point_cloud_res)
            point_cloud.copy_to(point_cloud_render)
            zed.retrieve_image(image_left, sl.VIEW.LEFT, sl.MEM.CPU, display_resolution)
            zed.get_position(cam_w_pose, sl.REFERENCE_FRAME.WORLD)

            # 3D rendering
            viewer.updateData(point_cloud_render, objects)
            # 2D rendering
            np.copyto(image_left_ocv, image_left.get_data())
            cv_viewer.render_2D(image_left_ocv, image_scale, objects, obj_param.enable_tracking)
            global_image = image_left_ocv
            # global_image = cv2.hconcat([image_left_ocv, image_track_ocv])
            # # Tracking view
            # track_view_generator.generate_view(objects, cam_w_pose, image_track_ocv, objects.is_tracked)

            cv2.imshow("ZED | 2D View and Birds View", global_image)
            key = cv2.waitKey(10)
            if key == 27:
                exit_signal = True
        else:
            exit_signal = True

    viewer.exit()
    exit_signal = True
    zed.close()


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='yolov8n.pt', help='model.pt path(s)')
    parser.add_argument('--svo', type=str, default=None, help='optional svo file')
    parser.add_argument('--img_size', type=int, default=416, help='inference size (pixels)')
    parser.add_argument('--conf_thres', type=float, default=0.4, help='object confidence threshold')
    opt = parser.parse_args()

    with torch.no_grad():
        main()

2.2 实验结果

测距图(感觉挺精准的)

视频展示:

Zed相机+YOLOv8目标检测跟踪

3. 版本二

3.1 相关代码

主代码 zed.py,具体放置在yolov8主目录下,可实现测距+跟踪+分割

python 复制代码
#!/usr/bin/env python3
import math
import sys
import numpy as np
from PIL import Image
import argparse
import torch
import cv2
import pyzed.sl as sl
from ultralytics.utils.plotting import Annotator, colors, save_one_box
from ultralytics import YOLO
from threading import Lock, Thread
from time import sleep

import ogl_viewer.viewer as gl
import cv_viewer.tracking_viewer as cv_viewer

zed = sl.Camera()

# Create a InitParameters object and set configuration parameters
init_params = sl.InitParameters()
init_params.camera_resolution = sl.RESOLUTION.HD720
init_params.coordinate_units = sl.UNIT.METER
init_params.depth_mode = sl.DEPTH_MODE.ULTRA  # QUALITY
init_params.coordinate_system = sl.COORDINATE_SYSTEM.RIGHT_HANDED_Y_UP
init_params.depth_maximum_distance = 20  # 设置最远距离

runtime_params = sl.RuntimeParameters()
status = zed.open(init_params)

if status != sl.ERROR_CODE.SUCCESS:
    print(repr(status))
    exit()

image_left_tmp = sl.Mat()
print("Initialized Camera")
positional_tracking_parameters = sl.PositionalTrackingParameters()
zed.enable_positional_tracking(positional_tracking_parameters)

obj_param = sl.ObjectDetectionParameters()
obj_param.detection_model = sl.DETECTION_MODEL.CUSTOM_BOX_OBJECTS
obj_param.enable_tracking = True
zed.enable_object_detection(obj_param)
objects = sl.Objects()
obj_runtime_param = sl.ObjectDetectionRuntimeParameters()

point_cloud_render = sl.Mat()
point_cloud = sl.Mat()
image_left = sl.Mat()
depth = sl.Mat()
# Utilities for 2D display
if __name__ == '__main__':
	model = YOLO("./yolov8n.pt")
    while True:
        if zed.grab(runtime_params) == sl.ERROR_CODE.SUCCESS:
            # -- Get the image
            zed.retrieve_image(image_left_tmp, sl.VIEW.LEFT)
            image_net = image_left_tmp.get_data()
            zed.retrieve_measure(depth, sl.MEASURE.DEPTH)
            zed.retrieve_measure(point_cloud, sl.MEASURE.XYZRGBA)
            img = cv2.cvtColor(image_net, cv2.COLOR_BGRA2BGR)

            result = model.predict(img, conf=0.5)
            annotated_frame = result[0].plot()
            boxes = result[0].boxes.xywh

            for i, box in enumerate(boxes):
                x_center, y_center, width, height = box.tolist()
                point_cloud_value = point_cloud.get_value(x_center, y_center)[1]
                point_cloud_value = point_cloud_value * -1000.00
                if point_cloud_value[2] > 0.00:
                    try:
                        point_cloud_value[0] = round(point_cloud_value[0])
                        point_cloud_value[1] = round(point_cloud_value[1])
                        point_cloud_value[2] = round(point_cloud_value[2])

                        distance = math.sqrt(
                            point_cloud_value[0] * point_cloud_value[0] + point_cloud_value[1] *
                            point_cloud_value[1] +
                            point_cloud_value[2] * point_cloud_value[2])
                        print(distance)

                        dis = []
                        dis.append(distance)
                        text = "dis:%0.2fm" % distance
                        cv2.putText(annotated_frame, text, (int(x_center), int(y_center)),
                                    cv2.FONT_ITALIC, 1.0, (0, 0, 255), 2)
                    except:
                        pass
                cv2.imshow('00', annotated_frame)
                key = cv2.waitKey(1)
                if key == 'q':
                    break

            zed.retrieve_objects(objects, obj_runtime_param)
            zed.retrieve_image(image_left, sl.VIEW.LEFT)
    zed.close()

3.2 实验结果

可实现测距、跟踪和分割功能,这个代码没有加多线程,速度够用懒得写了,对速率要求高的可以自己写一下,实现不同功能仅需修改以下代码,具体见 此篇文章

python 复制代码
 model = YOLO("./yolov8n.pt")
 img = cv2.cvtColor(image_net, cv2.COLOR_BGRA2BGR)
 result = model.predict(img, conf=0.5)

测距功能

跟踪功能

分割功能

视频展示

相关推荐
查士丁尼·绵30 分钟前
笔试-士兵过河
python
weixin_466832 分钟前
编程之python基础
开发语言·python
打酱油的;39 分钟前
【无标题】
爬虫·python·php
幸福清风1 小时前
【Python】基于Tkinter库实现文件夹拖拽与选择功能
windows·python·microsoft·tkinter
海琴烟Sunshine1 小时前
leetcode 168. Excel 表列名称 python
python·算法·leetcode
java1234_小锋2 小时前
TensorFlow2 Python深度学习 - 卷积神经网络示例2-使用Fashion MNIST识别时装示例
python·深度学习·tensorflow·tensorflow2
@sinner2 小时前
《扫雷:病毒蔓延》- 颠覆传统的动态扫雷游戏
python·游戏·pygame
愈努力俞幸运2 小时前
python 列表浅拷贝 深拷贝
python
测试老哥2 小时前
测试用例之正交试验法、功能图法
自动化测试·软件测试·python·功能测试·测试工具·职场和发展·测试用例
ARM+FPGA+AI工业主板定制专家3 小时前
基于JETSON+FPGA+GMSL相机 vs 传统工业相机:高动态范围与低延迟如何重塑机器感知视觉?
人工智能·数码相机·机器学习·自动驾驶