ฝัง
- เผยแพร่เมื่อ 14 มิ.ย. 2024
- The YOLOv8 team just released (March 9, 2023) native support for object tracking algorithms (ByteTrack and BoT-SORT): docs.ultralytics.com/tasks/tr...
The video will show you how to use the new YOLOv8 object tracking functionality using a live webcam stream.
Whether you're an experienced developer or a beginner in the field of computer vision, this video covers everything you need to know to get started with object tracking using YOLOv8. With step-by-step instructions, you'll learn how to use state-of-the-art deep-learning models to track objects in videos and live streams.
Chapters:
00:00 Introduction
00:44 Setup Python environment
01:28 Real-time object tracking in the terminal with YOLOv8 CLI
02:17 Bulk video tracking with YOLOv8 SDK
04:56 Setting up inference loop with YOLOv8 SDK
06:11 Bounding box annotation with Supervision
08:18 YOLOv8 tracking ⭐
10:26 Real-time object counting with Supervision
12:57 Conclusion
Resources:
⭐ GitHub repository with demo project: github.com/SkalskiP/yolov8-na...
🌏 Roboflow: roboflow.com
🌌 Roboflow Universe: universe.roboflow.com
⭐ Supervision repository: github.com/roboflow/supervision
🎬 Track & Count Objects using YOLOv8 ByteTrack & Supervision: • Track & Count Objects ...
🎬 Count People in Zone | 3 Models: YOLOv5, YOLOv8 and Detectron2: • Count People in Zone |...
🎬 YOLOv8 Object Counting in Real-time with Webcam, OpenCV, and Supervision: • YOLOv8 Object Counting...
🎬 YOLOv8: How to Train for Object Detection on a Custom Dataset: • YOLOv8: How to Train f...
📓 Learn more about YOLOv8 and other Computer Vision models with Roboflow Notebooks: github.com/roboflow/notebooks
Stay updated with the projects I'm working on at github.com/roboflow and github.com/SkalskiP! ⭐ - วิทยาศาสตร์และเทคโนโลยี
Thanks for keeping us update on new changes
I’m doing my best! 👍
Thanks you so much for the tutorial, this new feature is so nice!
I’m glad you like the tutorial! As for the feature, yup… it is great! 🔥 works out of the box, and doesn’t require a lot of changes in your code. It is pretty cool that you can build such a complicated app in just 30 lines of code!
Thanks you so much for the tutorial!!! Very cool review ! And THANKS again!
Thanks so much for the kind word ;)
Such a great video not only that it is very easy to understand also very usefull for projects. I hope you will continue to make videos like this, thank you!
Doing my best :) we already have few ideas in the pipeline.
Great job, thanks Piotr👏
As usual, thanks a lot for the support! 🙌
Thanks a lot lot lot. You r the best !!!
I needed so much to develop a photo analysis system, but I'm trying a professional and it's very difficult, in fact I've been trying for months.
Seeing you developing things, you could do what I need in one day, it's hard to do.
Hugs, congratulations for the work.
Oh, you shouldn't feel this way. What you see in that video is just me showing a project that I already have figured out. If you want to see the process of me learning how stuff works take a look here: th-cam.com/video/X9z6XaHMLZI/w-d-xo.html I spend over 2h putting that small script together.
thank you so much for a great video. 👍🏻x1000000000
Thank you very much!
Thanks for watching?
Thank you
Hey great video, I wanted to ask whether you could explain how you could allow a pan/tilt system to move the camera to the centre of a tracked class. Example, move the pan and tilt camera to the centre of a person to track them as they move. Any help would be greatly appreciated!
Is it possible to make the line into a circle zone? So an object is inside or outside a circular zone it gets counted based on that. Your videos are incredibly helpful. Thank you for the amazing content.
Great video! How can i extract the detections from the line to use it for example in a GUI, do some calcs, etc.? Thanks in advance!
hello sir! thank you for the tutorial, do you have a video on how we can extract the coordinates of the bounding box? i am trying to determine the centroid of the bounding box
As a more general discussion around the state of the state. What do you think is next for the task of re-identification? Which platforms are making progress in this regard? It's so important to re-identify people and objects and occlusion is a constant..
Thank you…
no entiendo porque esto me da error : labels = [
f"{model.model.names[class_id]} {confidence=0.2f}"
for _, confidence, class_id, _
in detections
]
hello , i have an issue when run the sript on Apple mac mini M1 chip, it takes around 300ms and about 3 frames per second and so laggy, so how can we add the device reconigsion so that the yoloV8 can leverage the apple silicon chip ?
Thank you
I followed the tutorial, it is great for the as is use case, but how do I import my own trained model from roboflow and use my own video, is there a documented process? Everything I try doesn't appear to be working
Hello, is it possible to recognise objects if they will appear in frame again, with same id? Like person, what if cloth will change?
How are you processing frames at 4ms? Would a NVIDIA Quadro P620 be powerful enough?
hi, is it possible to just use the tracker without the line counter?
Hello, how could I correct the code so that the file reads an input video and it exports an output video?
Hi!
Maybe I'll ask a stupid question... But my webcam is hidden, it's installed in a button.
when I want to turn on the camera, I have to press a button and at that moment the code crashes.
Help solve the problem. i think it has to do with setting cv2.imshow("yolov8", frame) or c2.waitKey(30)==(27)
i really need your help and hope you have run into this issue before. i am building a detection system on top of an existing (super old) security system for a farm. The screen i want to monitor shows 4 different security cameras at the same time, so i thought i could easily launch an OBS virtual cam of the whole screen and feed it into YOLOv8. No matter what i try, i can't get this virtual camera be read properly, it's always a blank grey screen. Inference from pictures or videos works brilliantly, virtual webcam not at all... i have scoured Github/StackOverflow etc for the past 3 days and i'm slowly loosing my mind. Do you have any experience with this? It seems to be a constant issue with multiple version of Yolo and OBS.
My alternative idea is to take screenshots via dxcam, mss or pyautogui, put i can't figure out how to feed them into yolo/ultralytics
Thanks for video, it's possible doing tracking from multi camera ?
You mean to track object across multiple cameras. It is but it would require much more complicated script to do that. It also can be much easier if camera views overlap each other.
such a great video ,i have a question is this project deployed on jetson nano?
Hi 👋! Thank you! No, it is not. I was running it on my GPU. But it should work on Nano just fine :)
I'm usinf the same git repository but instead I'm using it for person only. My in out counters remains 0 and they do not get updated
I wonder why I always see the issue " type object Detections has no attribute 'from_yolov8' " appear in every program I run :) Could you please help me?
Can you please help me in preparing a drone detection and tracking using webcam setup using your already existing models. Could you tell me how I can do it.i nees to do it using python so that i can later use it to control the pan and tilt of the camera mounted on a stand
i tried to make it on youtube video or video in the project path many errors appeared
can you help me
Great video. Just a question, how can i print/get the count in the code so that i can store in the database ?
You can access in_count and out_count values. Take a look here: supervision.roboflow.com/detection/tools/line_zone/
nvm. everything i did somehow didnt change the result. the code works but nothing has changed compared to when i casually use model.track without anything involved. its just weird.
I love your work and @Roboflow especially, I was trying to count multiple object at once, like if apple crosses the line, need to increase the count of apple likewise for all classes in video. output will be like apple 20, orange 11, banana 17 like this. need to get the count for all classes that we trained in custom model when it crosses the line, if possible please make a video or write an article that will be more helpful. thank you.
Hi 👋🏻 so per class line counting?
@@Roboflow let's say, the apple, orange and banana cross the line at same time, count of apple, orange and banana has to be incremented by 1, from apple 0 to apple 1, orange 0 to orange 1, banana 0 to banana 1, hope this makes my query more clear.
@@sathishk3494 understood. That feature request is something we get quite often in the comments on TH-cam. However no one yet created a feature request issue on GitHub. Would be great if you could do that: github.com/roboflow/supervision/issues
@@Roboflow hope i will do that, i was trying to do that, thanks for your fast reply, love your work, cheers!!!
@@sathishk3494 no worries! I try to do my best.
Great! Can you replace a webcam source with a video path? Is there a standard format for storing the annotations?
I’m happy you liked it! You mean dataset format to store tracking data?
@@Roboflow thanks for the response! Yes, for the tracking annotations. Using them is the next step once you can do tracking. I’m wondering if there is a common format for that type of data. The time series aspect of the annotations makes it so there could be many different ways to store it. But probably some are more efficient than others.
@@willemweertman1178 hi 👋🏻 this is the most commonly used dataset format for tracking I know: motchallenge.net/instructions/
No module named 'lap', I can't figure it out, does anyone got this error message?
so the realtime is based on the gpu we use right? I mean the speed of inference the yolov8 model, and I think also the FPS of the camera, so which gpu you are using and the webcam FPS?
I mean if you produce 30 fps stream, you want be able to run more than 30 fps sure. But usually you are primarily limited by model not camera.
This is basically just a channel for advertising ultralytics now.
Don’t worry! We are working on next video right now. It will have nothing to do with YOLOv5 or v8.
Hi. Great content. I have been working on developing the detection and tracking algorith and just upgraded to YoloV8. The problem is that YoloV8 does not use Cuda for my graphics card. i am using 3060 but only getting around 10 fps. Can you please help with that?
Did you installed gpu version of PyTorch?
@@Roboflow that must be the problem. thankyou so much. will try to install that.
@flyingnerds654 Would you have any idea what kind of graphics card I would need? Looking for something affordable like a NVIDIA Quadro P620.
How to set frame resolution to 720 to 1280 in this? l
dont know why if i use stream=true, it will show a paused version of the video.
Great Video. I am running the code on my Windows 11 laptop using the built in web camera and the code is working perfectly but the video feed is stuttering and there a big delay in the feed (almost 2 seconds). I am using and i7 processor 11th Gen and have 16 GB of RAM and I have a nvidia geforce mx450 display card. Any suggestions on how to fix this?
Do you have cpu or gpu version of PyTorch installed?
@@Roboflow It was running on the CPU. I used the n version of the yolo model and it became better. Thanks
is there something like box annotator for instance segmentation?
Not yet unfortunately. We Hase that on our TODO list. But I don’t have any specific date yet. I encourage you to create feature request here: github.com/roboflow/supervision/issues we will try to keep you updated.
On labels i have this error ValueError: too many values to unpack (expected 4)
i dont know to fix it and i only copy pasted the code
You are using a version of the supervision package that is too new. You need to downgrade your supervision to 0.3.0.
how do i track specfic object i know the index of it from coco dataset
Is there a native way to access last frame bounding boxes for a given track_id?
I don’t think so… you would need to build custom logic for that
Can you provide me what your setup is, please. I tested in several places (linux and windows) and always hangs in the part of yolo track.
Hi :) Did you used this repo: github.com/SkalskiP/yolov8-native-tracking ?
Thanks for the video, but can I somehow send videoCapture already?
This is really good question. I was wondering about the same thing. And to my knowledge you can’t. We have other tutorials where we use regular cv2 capture instead of the YOLOv8 video processing. So you can still do all those things. Just not with YOLOv8 tracking.
Hello, I'm having this issue with the lap module not being installed properly. You also seem to have this issue but it gets resolved in the "building wheel for collected packages". It's not getting solved for me and end up with the error "no module named 'lap'. Is there any solution for this?
Could you create a new issue here: github.com/roboflow/notebooks/issues ? Please make sure to provide as much detail as possible. I'll try to help you.
@@Roboflow that would be amazing thank you.
I am trying to apply the same concept in a production line with boxes passing through the detection line. The box has different prints on each side. So I need to track both sides of the box (2 cameras). I want to teach the model on one of the prints and if that object is detected it would count it. Is there a way to have multiple camera sources in the model.track or do I have to run the function two separate times or do you have any other recommendations? Thanks in advance
Just to clarify. The box passes through an opening and there is no control during production which orientation the box will be placed on the conveyor belt so the desired print to be detected could be on either side of the box.
@@alaaaladnani9693 sorry for late reply! I have a lot of work lately. So if I understand correctly you want to track those prints only not the whole boxes? And at the end you want to get count? Or what is the desired result?
@@Roboflow we have around 10 models (products) all boxes. We identify the product according to the print on one of the sides. The other side is a generic print that is available on all the boxes. So we trained our model according to the unique print. We also trained our model to identify this generic print.The unique print could be on either side of the box depending on the orientation of the box which we have no control over since it is done by humans working on the line. So we need cameras on both sides of the line to be able to identify the product. Currently we are running two simultanious instances of the tracker code one for each side using multiprocessing and it is working. My question is can we track both camera feeds on the same instace? Do you have any other suggestions?
@@alaaaladnani9693 now I understand! I think what you have is the optimal solution. Single tracker won't work. At least any of the trackers I know won't work if it gets coordinates from different cameras at once. I think if you try to merge it you will end up with more problems than benefits.
Will you be making a Colab notebook for this tutorial?
You think I should? There is a open source repository that you can use. Link is in the description bellow.
How do i store the output as a video instead of showing it directly in the window?
You can use supervision.roboflow.com/utils/video/#supervision.utils.video.VideoSink. Here is an example where I do it: github.com/roboflow/supervision/blob/develop/examples/traffic_analysis/script.py
Why can I do object detection with my CPU, but as soon as I select my GPU it gives me errors in regards to the model? Do you have any tutorials on how to get GPU to work with YOLOv8
What GPU do you have?
@@Roboflow I have a 2080ti that is compatible with CUDA
@@suaie interesting. What is the error you see?
Which version of python you're using for the application? I use 3.10 and lap module can't be installed.
Hi :) 3.9
@@Roboflow I'm using 3.9.13 and am having the same issue as Givaldo - the lap module doesn't install.
can you make a video with segmentation (ad example for changing background)
I absolutely love the idea. So kind of what Google Meet or Teams or Zoom are doing? Real time segmentation + background swap?
@@Roboflow Real time segmentation + background swap would be great, then if you use pyvirtualcam you can make it work for any external videoconf app. I did it wiith mediapipe, but it would be nice to do the same with yolo.
@@g.s.3389 this absolutely awesome idea! I’d love to try it out. We have few other things planned but I’ll keep that idea in mind. Thanks a lot!
is there possible to make a video of using torchreid with custom dataset?
Oh I actually haven’t used that project before. I’m not sure we will create tutorial for it. We usually try to stay away from projects that allow individual person recognition.
@@Roboflow it could be animals:) i am sure that Roboflow will one day have tools for working with more sophisticated tracking
@@gbo10001 I'm not saying no :) let me pass that through rest of the team
Thanks for sharing. I´m having an issue executing this script. > requirements: Command 'pip install "lap" ' returned non-zero exit status 1. Any tips for how can I solve this?
Oh. Didn’t expected something like that to happen. What OS are you on?
@@Roboflow Windows 11
@@wellyngtonalberti2002 looks like they support Windows 11. Did you tried to just do “pip install lap” before you do “pip install ultralytics”?
@@Roboflow I'm using Python 3.9.13 and am having the same issue. It looks like the current version of lap is using a deprecated install function. It says `numpy.distutils` is deprecated since NumPy 1.23.0, as a result of the deprecation of `distutils` itself.
Same issue
I want to know if you can develop crack detection in a masonry construction wall
I think you can. We have done crack detection on concrete wall some time ago.
"I congratulate you on a great video you made, but I need to see if I can go further with a project I'm working on. I see that you've drawn the object counting line, could we use line_counter.trigger(detections=detections) to record a detection data, such as xyxy=results[0].boxes.xyxy.cpu().numpy(), every time it passes through the line?"
Ski you would like to increase the counter every time that object passes the line? But you are still I tweezed in single counter or you would like to have separate counter for every tracker id?
@@Roboflow My model is yolov8_segmentation, with xywh I can calculate the size of the object I detect and save it in a variable that I then pass to box_annotator-labels as you do in the video, so far so good. What I want to try is that every time the detection counting line is passed, it saves those values that the labels had at that time, for example, to later send it to a database or present it in a pop-up window. And I was wondering if that can be done with the line_counter.trigger?
@@PabloGuillen1984 now I understand… unfortunately not. And I’ll be fully honest with you I doubt we will ever have that functionality. Looks like something really custom and specific. We try to facilitate reusable components.
Can you show how to use this in the colab environment? I need to change the code and get rid of bytrtrack, Thanks
The code itself should work. The difference is that you wont be able to use webcam. So you need to load some static video file.
Amazing stuff, that means no more Deep-SORT eh
If you use YOLOv8 you can run other trackers really easily! Awesome!
when I try to run the real time program code the camera doesn't show up? please help me 🙏
Hi 👋! Please give me more details. Do you see any errors in the console? What is your version of Python? What is the OS you use? What are the versions of pip packages that you have installed? Do you run it in docker?
@@Roboflow i use python version 3.9.16, OS windows 11, pip version 23.0.1, and so many pip packages in my env, and already installed supervision, ultralytics
why when run this program not count objek? still zero when i move objek
Could you give me a bit more info?
hi man, i failed to install 'lap modul' when i tried. i need solution. thank you
pip install lap does not solve your problem?
@@Roboflow this not solve my problem, when i do 'pip install lap', instead happened "Failed to build lap
ERROR: Could not build wheels for lap, which is required to install pyproject.toml-based projects"
do you have any solution?
@@Roboflow how much pip version?
@@alwishihab9845 I'd go for latest as possible
@@Roboflow when i run program, it tell me that lap not integrated to python and numpy update version now
thank you
how to use this with gpu ? with cuda
The only thing you need to do is make sure your PyTorch has GPU acceleration. YOLOv8 is as fast as your PyTorch is.
Plsss provide object detection and counting deployment model in web framework as a guide using bytetrack and supervision not only in terminal
Hi :) What do you mean by web framework? Could you be a bit more specific?
How do you get the number of people in and out
Do you want to access the values that are displayed on the frame?
how can we save the output as a file video?
Thank you by the way for the update!
@@user-no4ff6mq1r no worries. You can use supervision to do that :)
Can yolo v8 run in amd gpu?
This is a great question. I’m Reality it is not about YOLOv8 but PyTorch which is used under the hood. And since 1.8 they support ROCm that is AMD’s version of CUDA. In short I feel quite confident it will work.
@@Roboflow I read earlier from the official PyTorch website and the ROCm documentation, where it says it only uses Linux OS
@@LordWildbeast oh wow… I’m not the expert on AMD GPUs. And every time that I dive into this conversation I learn something new. Yes looks like it only runs on Linux.
how can i save this annotaion box to a target video?
do you want to save the result video?
@@Roboflow yes. With the box annotations.
@@gbo10001 did you tried using VideoSink: github.com/roboflow/supervision/blob/main/supervision/video.py
@@Roboflow i saw that you have it. But i am not sure how to use it:
Frame=....
What suppose to be inside?
@@gbo10001 numpy array with annotated frame of video
haha I thought the battery in my wired mouse is going to be used up soon
Thank for your helpful video! please give me source code count vehicle crossing the line. Thank you very much!
How can I get the original image captured from camera? I tried deepcopying and showing images from result.orig_img in the inference loop, but they seem to be annotated images.
Did you used numpy.copy?
@@Roboflow Issue solved, much obliged! I was actually implementing this project using Pyside, it turns out it's not the problem with the copied image array but the image signal emitters in Pyside, they were too fast when sending both predicted image signals and original image signals so they sort of displayed the mixed frames of annotated images and original images on the Pyside panel😂, but I got it fixed with time.sleep
@@christopherwu6246 I'm a bit worried when you say " got it fixed with time.sleep" but the most important thing is you managed to solve your problem.