Track & Count Objects using YOLOv8 ByteTrack & Supervision
Вставка
- Опубліковано 31 тра 2024
- Let's build together an application to track and count objects using Computer Vision. We used YOLOv8 for detection, ByteTrack for tracking, and the latest python library from Roboflow - Supervision for object counting.
Chapters:
0:00 Introduction
1:28 Setting up the Python environment for vehicle tracking
5:28 Using YOLOv8 for vehicle detection
6:27 Building custom inference pipeline with Supervision for a single image
12:37 Building custom inference pipeline with Supervision for a whole video
15:46 Tracking detections with ByteTrack
17:40 Counting objects crossing the line with Supervision
19:29 Training YOLOv8 Object Detection model on custom dataset
22:50 Detect, track, and count candies on the conveyor
25:50 Conclusion
Resources:
🌏 Roboflow: roboflow.com
🌌 Roboflow Universe: universe.roboflow.com
⭐ Supervision repository: github.com/roboflow/supervision
📝 Track and Count with YOLOv8 Blogpost: blog.roboflow.com/yolov8-trac...
📓 Track and Count Vehicles with YOLOv8 + ByteTRACK + Supervision Notebook: colab.research.google.com/git...
📓How to Train YOLOv8 Object Detection on a Custom Dataset Notebook: colab.research.google.com/git...
🎬 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...
🎬 Instance Segmentation in 12 minutes with YOLOv8 and Python: • Instance Segmentation ...
📓 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! ⭐ - Наука та технологія
your code comes with the error of numpy between float, int and double
I recommend using updated version of the notebook: colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-track-and-count-vehicles-with-yolov8-and-supervison.ipynb
thanks
@@Roboflow
@@Roboflow Does that notebook solve the error? Thank you!
Is there any other updated notebook because this one is also throwing some error
@@mariacaleropereira2967 The notebook works perfectly :)
The best and complete tutorial for implementing YOLOV8 based object detection, tracking and counting system. Love it brother
That’s what I strived for! Great to hear you liked it so much 🔥
How can I count the bounding boxes for a set of images ( not a video) in this case ( using a pre trained yolov8 model with only 1 class)
@@ashishreddy2634 Are you trying to detect specific class?
Always best contexts with very clear explanations... You are perfect bro !
haha! cv bro!
26 minutes for this is not long at all. Thank you for what you do and please don't hesitate to make longer videos, however you see fit.
My pleasure!
It goes to show how streamlined this stuff has become. Try doing a PhD in this ten years ago and having to write your own code for everything AND the novel parts you're working on. Takes months and hours to explain. Now anyone can git clone and run complex models. What a world :)
@@katanshin Truly!
Fantastic tutorial, playing around with plenty of the options here, thanks for the upload.
Hi it is Peter from the video 👋Thanks a lot! Let us know what other feature could be useful ;)
Thank you so much.. I have Zero experience on this matter but following each of your instruction and I did finish my project with my own video.. Super!
Did you do it locally?
Thank you brothers, for your work!
TY for your great work on supervision library. I have modified your line counting algorithm. During counting people from indoor cctv camera, lines stay short to meet counting conditions. Firstly, I tried center dot instead of corners of bb, but it become unstable, especially when a person pass from door, because center of rectangle become unstable while object slowly disappear. Finally, I draw a square at center of object. It fits my case and generate stable countings.
You deserve more subscribers and likes ! Cool guy and straightforward 💛
I hope we will get 50k subs this year! 🤞🏻
@@Roboflow Guys show your love for this dedicated Gentleman by subscribing and liking his content.
Just wow!
Thank you for this great content.
i love how u explain the tuitorial so awesome,thanks
This is so nice to hear. Make sure to try Supervision if you like the tutorial.
Awesome awesome awesome! Thank you for the excellent work
Great video! However I tried implementing it with more than one counter (one for each lane) but it seems that LineCounter is a global variable shared across all other lanes. is there a way to overcome this?
Thank you!
Hi, does the tracking work if the object changes class as time goes? For example for the football field dataset if my labels were "player static" and "player running" would the tracking be able to follow the same player as it changes states? Thank you!
really good staff!!!
Great one! very thorough well explained
Thanks a lot!
Thankyou so much. The explanation was in-depth.
My pleasure!
@@Roboflow by adjusting some resolution and having perfect line counter position,your code is doing great in real-time. 👍
Hello Piotr @roboflow, thank you for the video. I have trained my model on 3 different classes.Would it be possible to have the line zone annotator display the count of each class separately rather than the sum of detections of all classes? Can you please help with this?
Thank you for this tutorial, it helps us a lot
This is great to hear! 💜 I was hoping for such a positive feedback
I really enjoyed the last episodes, very well and comprehensibly explained! Thanks!
Would it be possible to make a video about rotated object detection in YOLOv8? Would be very useful.
Hi, it is Peter from video! Thanks from kind words. It means a lot to me. Is YOLOv8 capable of rotated object detection?
@@SkalskiP Hm you are probably right, rotated detection doesn't exist yet.
Thought I just overlooked it..
Thanks for the answer!
@@SkiLLFace360 no worries it is kind of my job to know it ;)
Great video tutorial. Thank you!!!
Thanks a lot 💜
hi. can i make a box instead of line, referencing on 17:40? so i want to count an object if that object staying in that box for milliseconds
Could you help with what each of the ByteTrackerArgs are for? I mean track_thresh, track_buffer and so on
You have touched my soul
Is there any easy way to count objects on pre predicted images? And print results in termina. I have a problem with find solution in internet.
how can i edit the in and out line (name/label), i want ti only detect in
and how i adjust the script to immedietely ipdate the count when the object enter the scanning line and so it wonr't recount the object that already ?
Hola tengo el error en la parte del código : tracks = byte_tracker.update( output_results=detections2boxes(detections=detections), img_info=frame.shape, img_size=frame.shape ) ;sale este error: AttributeError: module 'numpy' has no attribute 'float' ;pueden ayudar porfavor
Tengo el mismo error, pudiste solucionarlo?
@@manuelnavarrete4509 si ..antes de ejecutar el código agrega esta línea : !pip install -U numpy==1.23.5 ;después te pedirá reiniciar la sesión ,vuelves a ejecutar el código ya sin volver a instalar el numpy y listo
I am interested in running this algorithm in pycharm and with the YOLO model that I built and trained myself and not run the algorithm from a notebook.
Is it possible to get a customized py file?
For classification into car, bus, truck and motorcycle which one is used ByteTrack or Supervision?
Additionally, is Bytetrack also used for counting along with tracking? Because supervision is used for annotations.
hello, can i use the supervision to count object on yolov5? i have an existing onnx model
Thanks! I'm still a beginner, so this is very confusing. So I have one question: I can't use 'best.pt', the model trained with yolov7, in this tutorial, right?
How to check its performance because we always need accuracy, precision and other metrices to find the performance of the model ?, Do I need to annotate first then calculate the accuracy?
Can you please tell me how to calculate the overall accuracy(SHOWING mAP score) of the particular trained yolov7 model in Google colab?
Excellent video 👌
Thanks a lot! 🙏🏻 make sure to try Supervision!
thank you for the tutorial, very easy to understand! I have a question, how do I get the CSV file result to find out the coordinates of the bounding box?
so useful, thanks a lot
Hello how would I be able to print the values from the counter to a txt. file. I tried the following code but I am unable to obtain the count from the line_counter: with open("output.txt", "w") as file:
# Print the variable into the file
print(line_counter, file=file)
Classical... What a topic.. thanks..
No. Thank you for watching! ;)
@@SkalskiP As always amazing job! One problem I am facing inside match_detections_with_tracks function, when the object is not in frame and model return emtpy list this line gives error iou = box_iou_batch(tracks_boxes, detection_boxes)
How can I solve it?
@@shukkkursabzaliev1730 oh, that code is far from being bullet proof. Would you like me to update notebook to work for those use-cases?
Tengo problemas con numpy en la parte del tqdm, y ya cambie los np.float por float y aun asi el problema persiste
Great video, thanks a lot
My pleasure!
Is there any other updated notebook because this one is also throwing some error ?
Is there any change that I can make to detect the speed of the vehicles. I used two lines, and took a random value for the distance between those two lines. Now the problem is I am not able to understand how to use 'time' library to extract the time for the vehicles going up the lane and down the lane between the two lines. Can please anyone help me with that?
I have followed the code provided and the program runs well as shown in the video. If I want to count vehicles according to each class, such as the number of motorbikes and the number of cars. How and which parts should I change in the code? Thank you
this is awesome!
Hiw can i get the count of in and out vehicles in one var only . I just want the whole count . Is it possible??
thank you
very helpful
That's what I wanted to hear!
Nice job ! love from china❤
Hi, it is Peter from the video! Thanks a looot! Love from Poland.
Thank you very much, really appreciate! I applied to my custom video, it does not count correctly. I saw in your video it also does not count correctly, how we can improve it?
Hi I have other question. Is it possible in the moment when the byteTrack algorithm count the candy or the car or bus or any other object is possible that the algorithm classify per class the counting for example if 4 cars pass the line in the video show how many cars pass the lines and show other count for buses and other for motorcycles is it possible for bytetrack?
ByteTrack is only responsible for tracking. The counting is done by line logic. Right now it is possible but you would need a separate line for every class. You don’t need to draw all of those line. But you would need LineZone object to perform the actual counting
thanks for the video. I noticed that even with a clear view of all the vehicles, you still lose track of the truck and it gets a new id. Is there a way to limit the number of ids that the objects get so that this doesnt happen? for example you only have 4 possible labels during the video and the algorithm has to select the most likely label when tracking.
Is is possible to solve those issues. Or to at least make them less frequent. But potential solutions are usually strictly tied to use-case that you are trying to solve. In our case you can notice that those id changes are happening only when cars are still far away or when they are partially ocluded by this large metal object hanging over the left lane. Thats why I would propose to discard objects that are in top half of image and only take into account those that are in bottom half - closer.
Congratulations for the video, very useful. I have a question: If I want to take a v8 model to mobile without losing fps effectiveness and model effectiveness. How would you do? Thanks.
from 7:49, the notebook from the link in the description doesn't have those lines, so where can I copy them to paste? Thank you!
I just checked. The line definition is there.
@@Roboflow I'm sorry but I don't understand! Could you please reply me with the link?
Help me please! I has a error in the next code.
the problem is here:
tracks = byte_tracker.update(
output_results=detections2boxes(detections=detections),
img_info=frame.shape,
img_size=frame.shape
)
I dont know how i can fix it
I have been working with ByteTrack for a bit now, but I have struggled on evaluating its tracking performance do you know if it is possible to check tracking performance of the individual objects using something live MOT metrics?
Yes, it is possible but you would need to have annotated data.
what to do if i want to show the vehicle counts based on their class. like car in: 1, bus in: 2, car out : 5, bus out: 6
i want to only count the vehicule entering not exiting.is it possible?i
Great video, i have a question in this video the linecounter from supervision will increase when all line from prediction box is through the line. Can you change it so just from bottom line / top line??
Ask this question here: github.com/roboflow/supervision Describe what you want to do. We will do our best to help you.
appreciate the elaborate explanation. Can we tag each of those objects with unique id? like car1, car2 ...etc
Can i draw multiple lines to count objects at multiple locations ?
I really need help for one thing.
How can you show the specific number of cars and trucks that have gone in and out.
For example:
3 cars and 1 truck in and
5 cars and 1 truck out
We don't have a dedicated feature yet, but you can build a workaround solution. Create two separate line counters. Filter detections by class, to get car and truck detections and trigger one line counter with car detections and the other with truck detections.
@@Roboflow I will try that, thank you very much!!!
Hi Peter, thank you very very much for your video and explanation! It was incredibly helpful.
I have a question: for the counting why do we need the line? Since we have the tracking and each object has an id, should not be enough to count the unique id? Thank you
Can we have the car speed and the postion (x,y,z)
Hi I have question in this case you don't use deepsort technique for tracking the cars do I?
I use BytetTack. DeepSort is just another tracker that you can use.
Does the same code works for crowd videos ? I’ve been failing to do it.
Thanks.
It should. But I’d need to see specific result to understand what’s failing.
Hey there, thanks for the amazing YOLO 8 videos, I run the code for object detection and it was work fine. then I tried to run for instance segmentation. all steps are fine but in the final step when I run the code for Inference with Custom Model, code run without any issue but this message did not appear: Results saved to runs/segment/predict2. do you know what is the problem?
Could you create issue here: github.com/roboflow/notebooks/issues ?
@@Roboflow Hi I found the error , in the code should write : save=true but you forgot it I guess . Thanks
@@lofihavensongs thanks a lot! Let me try to update that
Hi I have a question what happen if in this case I would like to put a line in vertical way in the middel of the image how can I solve this problem because in your video you don't show how to obtain these values that you put for trace the line for counting
You just need to know your frame dimensions and experiment a bit to find the right fit.
thanks a lot this comment help me a lot @@Roboflow
i tried to use other video with your code but the for loop stops at 6% of the frame. it doesnt return error and it even runs next code properly. So what can be the problem?
SOURCE_VIDEO_PATH = "/content/highway.mp4"
video_info = VideoInfo.from_video_path(SOURCE_VIDEO_PATH)
byte_tracker = BYTETracker(BYTETrackerArgs())
TARGET_VIDEO_PATH = f"{HOME}/result.mp4"
generator = get_video_frames_generator(SOURCE_VIDEO_PATH, end=video_info.total_frames)
box_annotator = BoundingBoxAnnotator(thickness=4)
counts = []
ids = []
i = 0
print(video_info)
with VideoSink(TARGET_VIDEO_PATH, video_info) as sink:
for frame in tqdm(generator, total=video_info.total_frames):
print(frame)
i += 1
results = model(frame)[0]
detections = Detections(
xyxy=results.boxes.xyxy.cpu().numpy(),
confidence=results.boxes.conf.cpu().numpy(),
class_id=results.boxes.cls.cpu().numpy().astype(int)
)
tracks = byte_tracker.update(
output_results=detections2boxes(detections=detections),
img_info=frame.shape,
img_size=frame.shape
)
tracker_id = match_detections_with_tracks(detections=detections, tracks=tracks)
detections.tracker_id = np.array(tracker_id)
labels = [
f"#{tracker_id}"
for x
in detections
]
frame = box_annotator.annotate(scene=frame, detections=detections)
for x in tracker_id:
if not x in ids:
ids.append(x)
counts.append(len(ids))
sink.write_frame(frame)
print(i)
Piotr jest super-duper ultra yolo guru :D
It's Peter from video. I'm not sure if I'm YOLO guru, but thanks a lot for this kind comment. I went through a bit of internet hate lately, so it is great to here some positive feedback.
thank you, nicley done, I was wondering if we use the segmentation model, how can we annotate the segments with supervision ?
Great question. We have support for segmentation on our road map, but it will take us a bit more time to put it on production.
Thank you for the amazing video! Is it possible to invoke yolo8 on every 4th frame (for example), instead of every single frame? And have some kind of other system follow the object in the other 3 frames (to save on resources).
Not to my knowledge. You skip the frame completely or not. All of those trackers depend on boxes being generated by the model. That being said you can try to pass detections to tracker every 4th frame. It all depends on input video but could still work.
Is it possible do this on instance segmentation model?
i have this problem ''AttributeError: type object 'Detections' has no attribute 'from_ultralytics'''
i am new to this and collab so maybe thats why i am facing a small issue!
the code runs succefuly but output video is not showing..where did i go wrong?
the commands and all are exicuting with no issues and it even show its detecting cars in the final tab but no video
You need to download the result after processing.
@SkalskiP As always amazing job! One problem I am facing inside *match_detections_with_tracks* function, when the object is not in frame and model return _emtpy list_ this line gives error *iou = box_iou_batch(tracks_boxes, detection_boxes)*
How can I solve it?
Hi it's Peter from the video. I just fixed that problem. Could you try the tutorial once again?
even i upload the zip file it shows no directory .what can i do?
Thank you so much for the video. what's the difference between this notebook and using "yolo track model=path/to/best.pt tracker="bytetrack.yaml"" ?
Hi! That video was actually recorded before YOLOv8 team added tracking capability. But in short, you can use ByteTrack with any object detection model, and if you will use Ultralytics implementation then you are bound to use only YOLOv8.
Bro I'm getting problem whenever I'm installation supervision in g-drive
Please let me know how to solve this problem
hi bro thanks for your effort that tutorial is awsome , I need to do the same thing but real-time by using my webcam or an ip camera
hi! thanks ! its v useful. Can it be applied on cellphones ? like an android or IOS app?
I have one question: Since we are trying to count the objects and since the Object's id given by the tracker are unique, why can't we just count the last Id or count the different number of ids?
How do you know how many of them traveled up and how many down?
Suuperb... What if I want to detect and track the faulty chocolates in that video and mark the chocolate faulty until it leaves out the frame? Any thoughts on this?
Do you have a model to detect those faults?
@@Roboflow No, currently I have a model to detect potatoes on a conveyer belt. For detecting defects I'm thinking of using OpenCV to detect color deviations.
My problem is since potatoes keep rotating on the conveyer belt, I want to track the defective potato even if it keeps rolling.
Hey Peter!
Any thoughts on this? And also, Just now saw your video on Grounding DINO it looks interesting. What are your thoughts on using it to detect rotten/spoiled potatoes as explained in earlier comments.
@@snehitvaddi sorry I missed your comment. If you have images of rotten potatoes you can try if DINO detect it. Sounds like something that should work. Color range is doable as well, just pretty hard to get right color ranges I think :/
Thank you ,Could you please explain how to count objects detected in images?
I want to use supervision for face detection and tracking with Detectron2 model
How should I edit the code to draw multiple lines?
Thank you for a very good explanation.
I found that YOLOv8 has their own tracking command both CLI and python mode.
I tried on CLI mode, it works well. but unfortunatly in python mode, ID always reset to id 1.
Now I am thinking of using ByteTrack as you did or do you have any idea of using straight forward way to use YOLOv8 to tracking object.
Thanks,
Nott
Yeah we have video on YOLOv8 native tracking. Take a look here: ua-cam.com/video/Mi9iHFd0_Bo/v-deo.html
@@Roboflow thank you
Thanks for information. By the way, I could not understand about the BYTETracks classes.
In particular, I've tried adjusting the parameters aspect_ratio_thresh and min_box_area, but there wasn't a significant change. I would appreciate it if you could explain the specific meanings of these parameters. I hope to receive a prompt response. Thank you!
How do we count for each class
Is the code just related to one or two test cases/videos? Is it possible to do it for any video in general?
Oh! It should work for any video you want. I already seen so many projects build on top of that code demo. Let me know if that works for your case too!
Fun Fact: Tqdm is an arabic word pronounced "Ta-qa-dom", which means progress
Hi it's Peter from the video! Wow! I didn't know that. Now you made me look and here is what I found: tqdm derives from the Arabic word taqaddum (تقدّم) which can mean “progress,” and is an abbreviation for “I love you so much” in Spanish (te quiero demasiado).
@@SkalskiP didn't know about the Spanish abbreviation,
Nice informative tutorial btw
This is, in fact, fun. Thank you.
Difficult to install though no module ultralytics🙄
What if i want to count objects of a specific class only? how can this be implemented? Let's say that we want to count only "car" class and not "truck" also? Can this be done?
sure! you just need to filter detections before passing them through the line counter. You would need to add this line before triggering:
detections = detections[detections.class_id == YOUR_CLASS]
@@Roboflow well i get error: TypeError: 'Detections' object is not subscriptable
Maybe if i use this filtering before the line counter??
#class to filter
mask = np.array([class_id in CLASS_ID for class_id in detections.class_id], dtype=bool)
detections.filter(mask=mask, inplace=True)
Let's say instead of CLASS_ID to use [1] so i can keep only class 1
I cannot import lapjv. It show ImportError: cannot import name 'LARGE_' from 'lapjv' (C:\()\lapjv.cp38-win_amd64.pyd). I tried to solve this problem but cannot.
Thanks for your interesting video. Could you make a video to compare YOLOv8, YOLOv7, YOLOv6 for object detection and object tracking? That would be great!!!
Interesting idea! Do you think it is worth comparing them they are all super close regarding accuracy and speed. What sort of benchmark are you mostly interested in?
Great video! How do I customize the counter? For example, position it in the corner of the screen, count cars, trucks, and motorcycles with their own counters? Thank you!
did you find a fix?
Great video! how can we use gpu with yolov8? and TensorRT? thank you☺
How do I get the specific time stamp for which the object was early detected in the video
We don’t have time analysis support yet in supervision :/
Hello Piotr @roboflow , I want to be able to activate my previously trained intelligence so that it detects and counts in real time, very similar to the video, but without a video itself, how could I do it? :c
I’m afraid I don’t understand the question :/ could you help me out and give me more details?
Excellent! Just something I was looking! Thanks Roboflow.
What was the fps?
We hope you will build something cool using supervision pip package ;)
very helpful video, but my code kept on having error due to the line [ tracks = byte_tracker.update( ] saying "AttributeError: module 'numpy' has no attribute 'float'. " plus when I use the google colab link in the description, i ran the byetrack and it encounters the same error, eventhough i didnt change any code i left it be and it kept on having the same error, so is it an update issue? can you please try and run your code in google colab in the description you gave? Because its seriously not working when i didnt even change any code
I recommend using this updated version of the notebook: colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/how-to-track-and-count-vehicles-with-yolov8-and-supervison.ipynb
@Roboflow thx a lot, I'll be sure to try this tomorrow
@@Roboflow is there a lot of difference in two methods ?
How can i draw multiple lines to count ?
Thank you. This is nice. I have a requirement to create vehicle detection model with good accuracy. Is it possible to create for me and work for this
I get this error with supervision = 0.1.0. AttributeError: module 'supervision' has no attribute 'get_video_frames_generator'
How do you import it? from supervision.video.source import get_video_frames_generator? That was the API used in 0.1.0.