开始使用

使用 Ultralytics YOLO26 进行 VisionEye 视图对象映射

什么是 VisionEye 对象映射?#

Ultralytics YOLO26 VisionEye 让计算机能够识别并定位对象,模拟人眼精准观察的能力。借助此功能,计算机可以辨别并聚焦于特定对象,就像人眼从特定视角观察细节一样。

VisionEye object mapping with YOLO tracking

使用 YOLO26 映射对象#

VisionEye 会在画面中设定一个固定观察点,并从该点向每个跟踪对象绘制射线,让你直观了解从单一视角观察场景的效果。将 vision_point 设置为观察者的像素坐标,然后使用 Python API 或 CLI 对视频运行此解决方案。

使用 Ultralytics YOLO 进行 VisionEye 映射
# Monitor objects position with visioneye
yolo solutions visioneye show=True

# Pass a source video
yolo solutions visioneye source="path/to/video.mp4"

# Monitor the specific classes
yolo solutions visioneye classes="[0, 5]"

VisionEye 参数#

下面是 VisionEye 参数表:

参数类型默认值说明
modelstrNoneUltralytics YOLO 模型文件的路径。
vision_pointtuple[int, int](20, 20)视觉系统使用 VisionEye Solution 跟踪对象并绘制路径的位置。
line_widthint2解决方案绘制的框、关键点和计数的线条粗细。
verboseboolTrue启用逐帧日志,记录输入形状、类别计数和处理速度。跟踪调用本身始终不会输出日志。

你还可以在 VisionEye 解决方案中使用各种 track 参数:

参数类型默认值说明
trackerstr'tracktrack.yaml'指定要使用的跟踪算法。内置选项:botsort.yaml、bytetrack.yaml、ocsort.yaml、deepocsort.yaml、fasttrack.yaml、tracktrack.yaml。
conffloat0.25设置检测的置信度阈值;较低的值允许跟踪更多物体,但可能会包含误检。
ioufloat0.5设置用于筛除重叠检测结果的交并比 (IoU) 阈值。
classeslistNone按类别索引筛选结果。例如,classes=[0, 2, 3] 只跟踪指定类别。
devicestrNone指定推理设备(例如 cpu、cuda:0 或 0)。用户可以选择 CPU、特定 GPU 或其他计算设备来运行模型。
imgszint or tuple640目标推理图像尺寸:正方形图像使用整数,或使用 (height, width) 对。图像尺寸会调整为模型步幅的整数倍。
max_detint300跟踪前每帧保留的最大检测数量,用于限制解决方案在密集场景中处理的物体数量。
quantizeint or strNone推理精度:对于非 CPU 上运行的 PyTorch 和 TorchScript 模型,使用 16 (FP16) 或 32/unset (FP32) 进行选择;其他格式使用其模型文件和运行时所选的精度。此参数替代已弃用的 half 参数。

此外,还支持以下一些可视化参数:

参数类型默认值说明
showboolFalse如果 True,则在窗口中显示带标注的图像或视频。适用于在开发或测试期间即时查看结果。
show_confboolTrue在每个检测结果的标签旁显示置信度分数,帮助了解模型对该检测结果的确信程度。
show_labelsboolTrue在可视化输出中显示每个检测结果的标签,便于立即了解检测到的对象。

VisionEye 的工作原理#

VisionEye 会在画面中设定一个固定观察点,并从该点向检测到的对象绘制线条,从而模拟人类视觉从单一视角聚焦多个对象的方式。该解决方案使用对象跟踪在各帧中持续识别对象,并以可视化方式呈现观察者(观察点)与场景中对象之间的空间关系。

VisionEye 类中的 process 方法会执行几项关键操作:

  1. 从输入图像中提取轨迹(边界框、类别和轨迹 ID)
  2. 创建标注器以绘制边界框和标签
  3. 为每个检测到的对象绘制框标签,并从观察点绘制视觉连线
  4. 返回带有跟踪统计信息的标注图像

这种方法尤其适用于需要空间感知和对象关系可视化的应用,例如监控系统、自主导航和交互式装置。

VisionEye 的应用#

VisionEye 对象映射在各行各业都有许多实际应用:

  • 安防与监控:从固定摄像机位置监控多个关注对象
  • 零售分析:跟踪顾客相对于店内陈列的移动模式
  • 体育分析:从教练视角分析运动员的位置和移动
  • 自动驾驶汽车:可视化车辆如何“观察”并优先处理周围环境中的对象
  • 人机交互:创建能根据空间关系做出响应、更加直观的界面

将 VisionEye 与其他 Ultralytics 解决方案(例如距离计算或速度估算)结合使用,你可以构建全面的系统,不仅能跟踪对象,还能理解对象之间的空间关系和行为。

常见问题#

  • 要开始使用 Ultralytics YOLO26 的 VisionEye 对象映射,先通过 pip 安装 Ultralytics 包。然后,你可以使用文档中的示例代码,通过 VisionEye 设置目标检测。下面是一个简单示例,助你快速上手:

    import cv2
    
    from ultralytics import solutions
    
    cap = cv2.VideoCapture("path/to/video.mp4")
    assert cap.isOpened(), "Error reading video file"
    
    # 视频写入器
    w, h, fps = (int(cap.get(x)) for x in (cv2.CAP_PROP_FRAME_WIDTH, cv2.CAP_PROP_FRAME_HEIGHT, cv2.CAP_PROP_FPS))
    video_writer = cv2.VideoWriter("vision-eye-mapping.avi", cv2.VideoWriter_fourcc(*"mp4v"), fps, (w, h))
    
    # 初始化 vision eye 对象
    visioneye = solutions.VisionEye(
        show=True,  # 显示输出
        model="yolo26n.pt",  # 使用 Ultralytics 支持的任意模型,例如 yolo26s.pt
        classes=[0, 2],  # 为特定类别生成 visioneye 视图
    )
    
    # 处理视频
    while cap.isOpened():
        success, im0 = cap.read()
    
        if not success:
            print("Video frame is empty or video processing has been successfully completed.")
            break
    
        results = visioneye(im0)
    
        print(results)  # 访问输出
    
        video_writer.write(results.plot_im)  # 写入视频文件
    
    cap.release()
    video_writer.release()
    cv2.destroyAllWindows()  # 关闭所有已打开的窗口
  • Ultralytics YOLO26 兼具速度、准确率和易集成性,非常适合用于对象映射和跟踪。主要优势包括:

    1. 实时性能:实时目标检测准确率高。
    2. 灵活性:在同一流水线中支持检测、跟踪和距离计算。
    3. 社区与支持:提供详尽的文档和活跃的 GitHub 社区,方便排查问题和改进功能。
    4. 易于使用:直观的 API 简化了复杂任务,让你能够快速部署和迭代。

    如需详细了解应用和优势,请参阅 Ultralytics YOLO26 文档。

  • Ultralytics YOLO26 可与 Comet、ClearML 等各种机器学习工具集成,实现实验跟踪、协作和可复现性。按照详细指南开始操作:将 YOLO26 与 Comet 集成和将 YOLO26 与 ClearML 集成。

    如需进一步探索并查看集成示例,请参阅我们的 Ultralytics 集成指南。

评论60

提出问题、分享结果,或帮助其他人。

  1. TranDHoaGitHub

    Hey, you can adjust the viewing angle of the eyes and make it measure the distance from the center of an object instead of from the edge.

    1. Glenn JocherUltralytics

      VisionEye already draws the ray from vision_point to the bounding-box centroid (not the edge) via SolutionAnnotator.visioneye() as shown in the VisionEye guide and the SolutionAnnotator visioneye() reference; there isn’t a separate “viewing angle” parameter today since it’s a 2D overlay, so the intended control is moving vision_point, and if you need an actual distance value (vs just a line), please use the Distance Calculation solution and share your desired behavior here (plus a short clip/frame + vision_point) so the team can advise the cleanest extension.

    1. Glenn JocherUltralytics

      @dixroby the requirements.txt file typically includes dependencies like torch, opencv-python, and other libraries necessary for running Ultralytics YOLO models. For a detailed list, please check the file directly in the repository.

    1. Glenn JocherUltralytics(已编辑)

      @FarrelCulbert VisionEye Object Mapping is not limited to YOLOv8: it works with the tracking-capable Ultralytics models you pass via model=, including YOLO26 (recommended) and YOLO11. For more detailed information and examples, you can refer to the VisionEye documentation at https://docs.ultralytics.com/guides/vision-eye/.

  2. skp4595GitHub

    rather than taking video as input fro testing, I want to test for an image, how can I do that? guide me on this.

    1. Glenn JocherUltralytics(已编辑)

      Hi there!

      Great question! If you want to test an image instead of a video using Ultralytics YOLO, it's quite straightforward. Below is a simple example to guide you through the process:

      Testing on an Image

      First, ensure you have the latest versions of torch and ultralytics installed. If not, please update them:

      pip install --upgrade torch ultralytics

      Next, you can use the following code to test an image:

      import cv2
      from ultralytics import YOLO
      from ultralytics.utils.plotting import Annotator, colors
      
      # Load the model
      model = YOLO("yolo26n.pt")
      
      # Load the image
      image_path = "path/to/your/image.jpg"
      im0 = cv2.imread(image_path)
      
      # Perform prediction
      results = model.predict(im0)
      boxes = results[0].boxes.xyxy.cpu()
      clss = results[0].boxes.cls.cpu().tolist()
      
      # Annotate the image
      annotator = Annotator(im0, line_width=2)
      names = model.model.names
      
      for box, cls in zip(boxes, clss):
          annotator.box_label(box, label=names[int(cls)], color=colors(int(cls)))
      
      # Display the annotated image
      cv2.imshow("YOLO Image Prediction", im0)
      cv2.waitKey(0)
      cv2.destroyAllWindows()
      Explanation
      1. Load the Model: We load the YOLO model.
      2. Load the Image: Read the image using OpenCV.
      3. Perform Prediction: Use the model to predict objects in the image.
      4. Annotate the Image: Draw bounding boxes and labels on the image.
      5. Display the Image: Show the annotated image using OpenCV.

      This should help you get started with testing images. If you encounter any issues or have further questions, feel free to ask! 😊

      For more detailed information, you can refer to the VisionEye Object Mapping Guide.

      Happy coding! 🚀

  3. skp4595GitHub

    I want to run VisionEye Object Mapping using YOLOv8 code on KITTI data set. the KITTI dataset is labeled as <class_name1, truncation, occlusion, observation angle alpha, bbox cordinantes, Dimension of object, Location, rotation_y, score > for each image. While model training, I'm getting this error train: WARNING ⚠️ /content/dataset/train/images/005936.png: ignoring corrupt image/label: could not convert string to float: 'Car' train: WARNING ⚠️ /content/dataset/train/images/005937.png: ignoring corrupt image/label: could not convert string to float: 'Pedestrian' train: WARNING ⚠️ /content/dataset/train/images/005938.png: ignoring corrupt image/label: could not convert string to float: 'Truck'

    with these args { EPOCHS = 5 !yolo
    task=detect
    mode=train
    model=yolov8s.pt
    imgsz=640
    data=depth_estimation_using_YOLOv8.yaml
    epochs={EPOCHS}
    batch=16
    name=yolov8s
    exist_ok=True
    amp=False } how do I fix this?

    1. Paula DerrengerUltralytics(已编辑)

      Hi! The model is trained using the YOLOv8 framework. You can train it on custom datasets like KITTI by converting the dataset to the YOLO format. The dataset should have labels in the format <class_id> <x_center> <y_center> <width> <height>.

      For more details on VisionEye Object Mapping, check out the Ultralytics VisionEye guide. If you have any specific questions, feel free to ask! 🚀

      Here's a sample training command:

      yolo task=detect mode=train model=yolo26s.pt data=your_dataset.yaml epochs=5 batch=16

      Happy training! 😊

      1
  4. skp4595GitHub

    what does this mean "It calculates this distance by analyzing the size and position of the object within the frame, using the camera's field of view and other calibration parameters if available"? what do you mean by 'other calibration parameter if possible'? The model is not a blackbox. Is it? there must be some calibration, some ground truth from where model must have learned. please explain this.

    1. Glenn JocherUltralytics(已编辑)

      Hello! Great questions about the VisionEye Object Mapping using Ultralytics YOLOv8.

      1. Distance Calculation: The phrase "It calculates this distance by analyzing the size and position of the object within the frame, using the camera's field of view and other calibration parameters if available" was inaccurate. The VisionEye distance example does not estimate camera-to-object distance from object size or field of view. It computes the pixel distance between the vision point and the bounding-box centroid and divides it by the pixel_per_meter value you set, so you calibrate that scale yourself.

      2. Other Calibration Parameters: The example uses none of these. You would need data such as camera intrinsic parameters (focal length, optical centers), extrinsic parameters (orientation and position of the camera), object size references, or a depth sensor only if you build a true camera-to-object distance estimate yourself.

      3. Model Transparency: YOLOv8 models are not black boxes in the sense that their architecture and training processes are open for examination and modification. They learn from labeled data (ground truth) where objects are manually annotated in images, teaching the model how to recognize similar objects in new images.

      For further details or specific implementations, feel free to check out our guide on VisionEye Object Mapping at Ultralytics VisionEye Guide. If you have more questions or need clarification, don't hesitate to reach out! 😊

    1. Paula DerrengerUltralytics(已编辑)

      Hey there! 😊 It sounds like you're encountering some version compatibility issues. For Ultralytics YOLOv8, we generally recommend using Python 3.8 or newer as it strikes a good balance between being up-to-date and maintaining compatibility with most dependencies.

      If you're using a Python version older than 3.7, you might run into issues, and it's a good idea to update to at least version 3.7 or newer if possible. Here's a quick way to check your current Python version from your terminal:

      python --version
      # or if you have Python 3 installed as 'python3'
      python3 --version

      And if you need to update Python, the official Python Downloads page is a great place to start.

      If you're still facing issues after checking your Python version, could you please provide more details about the error messages you're seeing? This information can help us pinpoint the problem more accurately. Thanks!

    2. Paula DerrengerUltralytics

      @mirhasnainsiba hi there! 👋 It seems there might be a discrepancy with the distance calculation in your setup. Typically, the distance calculated by the model would depend on the scale set in your system (for example, pixel_per_meter in the example). If the object is being reported as 61m away while it's actually about 1m away, you might want to adjust this scale factor accordingly. Here's a brief example:

      pixel_per_meter = 100  # Adjust this value based on your actual scale

      This adjustment helps calibrate the distance calculation to more accurately reflect real-world distances. If you continue to experience issues, double-check the calibration process or consider revisiting the portion of your code responsible for calculating distances. If further assistance is needed, feel free to reach out! 🛠️

    3. Paula DerrengerUltralytics

      Hey there! 😊 Thanks for reaching out. If you're observing an object distance as 61m while it's almost 1m away, you might want to adjust the pixel_per_meter scaling factor in your VisionEye implementation. This adjustment is crucial for accurate distance calculation.

      Here's a simplified approach:

      pixel_per_meter = 100 # This is just an example value. 
      # You'll need to calibrate this based on your specific setup and camera properties.

      By fine-tuning this variable, you can achieve more accurate distance measurements. Calibration involves comparing known distances with the pixel distances detected by your model to find the most suitable pixel_per_meter value. 🛠️

      Hope this helps! Let us know if you have any further questions or if there are more details you'd like to dive into.

    1. Paula DerrengerUltralytics

      Hey there! 🌟

      The video files used in our examples are not provided in the repository. You can use any video file that you have access to by placing it in an accessible directory and then modifying the "path/to/video/file.mp4" placeholder in the code snippets to point to your video file's location.

      For example:

      cap = cv2.VideoCapture("your/video/path/here.mp4")

      Just replace "your/video/path/here.mp4" with the actual path to your video. If you're looking for sample videos, online resources like Pexels Video or Pixabay Videos offer free footage that you might find useful for your projects.

      If you have any more questions or need further assistance, don't hesitate to reach out. Happy coding! 🚀

      1
    2. Paula DerrengerUltralytics

      Hey! Thank you for checking out the VisionEye Object Mapping with Ultralytics YOLOv8. 🚀 If you're keen to dive into VisionEye Object Mapping for pinpoint precision or even integrate object tracking and distance calculation, the provided Python examples can guide you seamlessly through the process.

      Just remember to replace "path/to/video/file.mp4" with the specific path to your video file, and you’re all set to explore the amazing capabilities of YOLOv8 with VisionEye!

      For pinpoint object mapping:

      model = YOLO("yolov8n.pt")  # Adjust model as needed
      cap = cv2.VideoCapture("your/video/path.mp4")  # Your video path here

      For any additional queries or discussions, feel free to drop a message in our GitHub Issue Section. Happy coding! 🎉

  5. AdadAlShababGitHub

    Hi, I just wanted to know about "VisionEye View with Distance". What does this 'Distance' means? I mean does it measure the distance from the camera to object? Please let me know. TIA.

    1. Paula DerrengerUltralytics(已编辑)

      @AdadAlShabab hey there! 👋

      No, the "Distance" in the VisionEye example is not a camera-to-object distance. It is the pixel distance between the vision point and the detected object's bounding-box centroid, divided by a pixel_per_meter scale that you set and calibrate yourself. Here's a simplified example of how you might calculate distance in a custom script:

      import cv2
      import math
      from ultralytics import YOLO
      
      model = YOLO("yolo26s.pt")
      cap = cv2.VideoCapture("path/to/video.mp4")
      
      while True:
          ret, im0 = cap.read()
          if not ret:
              break
      
          results = model.track(im0, persist=True)
          for box in results[0].boxes.xyxy.cpu():
              # Example: Calculate distance based on object size or position here
              # This is a placeholder for actual distance calculation logic
              distance = calculate_distance(box)
              print(f"Distance: {distance} meters")

      In this example, calculate_distance would be a function you define based on your specific distance estimation method. Keep in mind, accurate distance measurement might require camera calibration and knowing the actual size of the object or using depth sensors in conjunction with the camera.

      Hope this helps! Let us know if you have any more questions. 😊

    2. AdadAlShababGitHub

      Hi @pderrenger. Thanks a lot for your reply. I just wanted to know how can I divide the results by 10 ( ex: distance/10)? I'm dropping the code, please let me know in which line/lines I should make the changes? TIA.

      Code:

      if results[0].boxes.id is not None:
          track_ids = results[0].boxes.id.int().cpu().tolist()
      
          for box, track_id in zip(boxes, track_ids):
              annotator.box_label(box, label=str(track_id), color=bbox_clr)
              annotator.visioneye(box, center_point)
      
              x1, y1 = int((box[0] + box[2]) // 2), int((box[1] + box[3]) // 2)    # Bounding box centroid
      
              distance = (math.sqrt((x1 - center_point[0]) ** 2 + (y1 - center_point[1]) ** 2))/pixel_per_meter
      
              text_size, _ = cv2.getTextSize(f"Distance: {distance:.2f} m", cv2.FONT_HERSHEY_SIMPLEX,1.2, 3)
              cv2.rectangle(im0, (x1, y1 - text_size[1] - 10),(x1 + text_size[0] + 10, y1), txt_background, -1)
              cv2.putText(im0, f"Distance: {distance:.2f} m",(x1, y1 - 5), cv2.FONT_HERSHEY_SIMPLEX, 1.2,txt_color, 3)
    3. Paula DerrengerUltralytics

      Hi @AdadAlShabab! 👋

      To divide the calculated distance by 10, you can adjust the line where distance is calculated. Here's how you can modify it:

      distance = (math.sqrt((x1 - center_point[0]) ** 2 + (y1 - center_point[1]) ** 2)) / pixel_per_meter / 10

      This change will effectively scale down the calculated distance by a factor of 10. Just add / 10 at the end of the existing distance calculation line in your code.

      Hope this helps! Let me know if there's anything else you need. 😊

    4. skp4595GitHub

      what does this mean "It calculates this distance by analyzing the size and position of the object within the frame, using the camera's field of view and other calibration parameters if available"? what do you mean by 'other calibration parameter if possible'? The model is not a blackbox. Is it? there must be some calibration, some ground truth from where model must have learned. please explain this.

    5. skp4595GitHub

      why are we dividing "distance " by 'pixel_per_meter' ? what is the logic or explaination for this ?

    6. Paula DerrengerUltralytics

      Hey there! 👋

      The division by pixel_per_meter is used to convert the measured pixel distance into real-world units like meters. This conversion factor helps translate the scale of the image to actual distances based on how many pixels represent one meter in the image. It's essential for applications where you need to know the physical separation between objects or from the camera to an object, not just the distance in pixels.

      If you have more questions or need further clarification, feel free to ask! 😊

    7. skp4595GitHub

      you mean to say, if there's 1 unit Euclidean distance between two pixels in an image, than it'd be 10m in real world scenarios? And this scale can be adjusted according to our will as per "pixel_per_meter = 10" this code. Isn't it?

      if so, how is this calibration done? where's the ground truth?

    8. Paula DerrengerUltralytics

      Hey there! 👋

      Yes, exactly! If the pixel_per_meter is set to 10, it means that 1 pixel in the image corresponds to 0.1 meters in the real world. You can adjust this scale to fit the specific context or calibration of your camera setup.

      Calibration is typically done by measuring known distances in the real world and adjusting the pixel_per_meter value so that the distances in the image accurately reflect these measurements. Ground truth data for calibration might come from physical measurements or known dimensions within the scene.

      If you have any more questions or need further clarification, feel free to ask! 😊

    9. Paula DerrengerUltralytics(已编辑)

      You're welcome! 😊

      The "VisionEye View with Distance" feature calculates the pixel distance between a reference point and the detected object's centroid, not a true camera-to-object distance. The pixel_per_meter value is used to convert pixel measurements to real-world distances. Calibration involves setting this value based on known distances in your scene.

      To divide the distance by 10, modify the distance calculation line like this:

      distance = (math.sqrt((x1 - center_point[0]) ** 2 + (y1 - center_point[1]) ** 2)) / pixel_per_meter / 10

      For more details, check out the VisionEye guide.

      1
  6. leongjy2005GitHub

    Traceback (most recent call last): File "d:\Random_YOLO8_Tutorials\Human_Detection\visioneye.py", line 28, in annotator.visioneye(box, center_point) ^^^^^^^^^^^^^^^^^^^ AttributeError: 'Annotator' object has no attribute 'visioneye'

    1. Glenn JocherUltralytics(已编辑)

      @leongjy2005 visioneye is not a method of the base Annotator; it lives on SolutionAnnotator. Use from ultralytics.solutions.solutions import SolutionAnnotator and annotator = SolutionAnnotator(im0, line_width=2), or use the solutions.VisionEye solution from the VisionEye guide.

      1