Great video, really helpful. I am building a people tracker to count how many people are in various zones of a premises and where they have walked from so this is a really useful piece of work - many thanks
I personally prefer this format of video where you work from the IDE rather than Colab as I find it easier to follow the structure of your thinking, how you go about solving the problem piece by piece.
Hello Piotr, I hope you are doing fine! I'm at 11:49 min and I'm getting this error: /python3.9/site-packages/supervision/detection/core.py", line 28, in _validate_mask raise ValueError("mask must be 3d np.ndarray with (n, H, W) shape") ValueError: mask must be 3d np.ndarray with (n, H, W) shape I haven't been able to solve it. I also clone the repo and I got the same problem. 😥
thank you very much for the video and I asked if I wanted to analyze the congestion by determining the length of the traffic jams, the number of cars in traffic jam for each lane of traffic. THANKS!
A really great tutorial. I have one question though, while running script.py on your video it works fine, but when I run it on my video at 12% it quits with error " in _get_ufunc_and_otypes raise ValueError('cannot call `vectorize` on size 0 inputs ' ValueError: cannot call `vectorize` on size 0 inputs unless `otypes` is set". Any ideas why?(I added otypes=[object] as second vectorize parameter and I am waiting for 12 hours now.) And how could I display class name alongside tracker id and group the tracking count displayed on zones according to vehicle class?
I am working on a project, for detecting saplings in a field. but when the drone is taking a turn to change the lap (basically moving into other lanes) the previous lanes are also coming like 1,2 of them and saplings in those laps are counted again. Is there any solutions for this?
Did you tried any image stitching? So I’m short instead of prow easing video. Pick frames every 1 or 2 seconds. Stich them together to form large image. And process that large image?
@@Roboflow actually I have a video from client, I have yolov8 model to detect saplings and the video is shot from drone like if the drone move in first lap, it covers first 4 lanes and when takes a right U turn and move to the 2nd lap the video has previous last 2 or 1 lanes, and saplings on those lanes are counted again, will image stitching work?
Can you please upload a tutorial on using YOLO-NAS and BYTETrack for object detection and tracking on custom dataset? (if possible, pls include the annotation part for videos of the dataset as well)
I got a error while compilation with visual code where it state that objects are nun cannot iterate the null objects plz suggest how to solve that issue
can you make videos on this topic Traffic flow tracking at an intersection based on direction-to-direction analysis involves detecting vehicles and determining their movement direction within predefined regions
From which direction and how many vehicles have gone on signalized intersection with their class name ?" cannot be showed .. how can we place a name in the line and for example D is the one direction A is the another direction how can we show how manh vehicle can pass from d-A
Thank you for a great work! Would be happy to see you continue experimenting more😊 Question though, I assume Polygon instance only contains information whether the ID is in the zone or outside without actually showing whether it crosses the line to exit? The reason I’m asking this is that my projects has a lot of occlusion, so it’s quite often for model to lose track. What will happen if object is not tracked anymore, but it never did actually cross the line to exit? Thank you very much!
So what you are looking for is information that given ID was outside and at some point it cross and is inside? To prevent ID popping out of nowhere in the middle of the zone and get counted in?
Not exactly. I want be sure that actual zone exit (crossing the line) won’t be confused with model losing the object. Say the object disappeared (due to occlusion) inside the zone. Would the polygon logic treat it as zone exit?
@@Roboflow I'm eager to use and promote your framework. please let me know if you can provide this information: I want to be sure that actual zone exit (crossing the line) won’t be confused with model losing the object. Say the object disappeared (due to occlusion) inside the zone. Would the polygon logic treat it as zone exit?
Impressive job! I have one question, although this video was captured by a drone, the camera lens remains stationary. In most cases, the drone is moving, so fixed-position bounding box is not applicable. Is there a way to address this issue? Thank you!
I have lost hope in finding the answer of my question @roboflow team. I have been commenting on your last 3-4 videos for one help. I want you to help me with Autodistill auto annotations video,where I want to check the annotated images manually so that i can correct the annotation which is wrong and missing. You told me that you will upload a blog, if its already uploaded please share the link or please make a tutorial on the steps . Thank you .Please i need your help
thank for great tutorial. I want to implement a tracker which counts and track fruits on tree. The problem is some of the fruits are counted more than once. does anybody have any ideas? is there a way to keep the tracklets in ByteTrack for longer duration?
Thank you so much for this video, it is a lot fun learning from your videos! I have one question, how do I convert the results I have gotten into a output video? I have tried, however I have only gotten the video without any labels or bounding boxes.
oh I GOT IT! Thank you nonetheless for making these awesome videos! Is it possible for you to create a video whereby ReId is combined with tracking algorithm(eg ByteTrack or BOTsort)? Thank you once again!
Video from you really helps me a lot I appreciate it . I have a query How can I count number of objects entered and pass from polygon and show the count on polygon
impressive! Actually it would be nice to draw the zones (bounding box) using mouse dynamically and assign a color, how can that be done? any specific libraries for it?
For "solar panel counting using SAM" ...which process is better >>> 1. detection followed by segmentarion Or 2. segmentation followed by classification
Really great project! I would like to save the tracking data for further analysis - speed calculation, trajectory conflict points. What would be the best way to do it? Let's say the result should be a dictionary with all tracked IDs as keys and the values will be a list with the object coordinates in each frame: {id1: [[x1,y1,frame1],[x2,y2,frame2].......} Thank you!
I trained our network with custom dataset. with YOLOv8 The training period lasted a long time. I want to test the performance of my test set with the network I trained at another time. Is there any other solution than retraining the network?
We can do many different things. You can label with single or multiple categories. You can assign tags. You can cluster images based on similarity. Could you be a bit more specific. What solution are you looking for?
@@Roboflow thank you! reverse image matching to find the closest match from an image set using the entire image instead of labeling things inside the image... let me know!
Hey can you create a tutorial for smart Surveillance which detect human intrusion classify between authorised and unauthorised intruders and notify security person usign YOLO V8 please
I'd love to see how to make a system that tracks people on a, for example, clothing store, and generates a heat map of the most popular spots in the store O.o
Hey dude, it's going to sound a bit lame, but my training has failed many times already and I think it's because of my dataset, I need it to be able to detect cars in that same kind of camera angle, so... can you give that database or tell the link where you got it from please 😅
@Roboflow, all: Has anyone managed to run YOLOv8 on a Coral Edge TPU? I've converted the YOLOv8 model: model = YOLO('yolov8n.pt') model.export(format='edgetpu',int8=True) I get: Model successfully compiled but not all operations are supported by the Edge TPU. A percentage of the model will instead run on the CPU, which is slower. Number of operations that will run on Edge TPU: 230 Number of operations that will run on CPU: 26 Running the model with the "yolov8n_full_integer_quant_edgetpu.tflite" file (and a properly plugged in and recognized Coral Edge TPU USB) results in detection times at least two times slower compared to running on my CPU. - Are these the correct steps? - Is this an inherent problem of the YOLOv8 model? i.e. not all of it can be efficiently quantized to run on a Coral Edge TPU? - What hardware do you guys use for edge real time YOLO?
i personally enjoyed the IDE more than Colab and please do the ones you talked about. traffic density, vehicle speed even the gap between vehicles travelling in the same direction. This will be interesting. What if the zones can be entered from the UI instead of hard coding them. just a thought!
Hey Piotr, Huge fan of you but, I have one problem is I'm working in Windows Pycharm. And I have no clue how to run the code in mine ... When i run the code the below part of the code says rebuild it in Windows so if it's possible can you help me in this. Thanks for your work and you keep me motivated to learn more about CV !! Huge fan mate cv2.imshow("frame", processed_frame) if cv2.waitKey(1) & 0xFF == ord("q"): break cv2.destroyAllWindows()
Excellent work with great applicability to many fields.
Agree!
Great video, really helpful. I am building a people tracker to count how many people are in various zones of a premises and where they have walked from so this is a really useful piece of work - many thanks
Hi 👋🏻 I guild something similar in the past. Awesome! I’m happy to hear that. I hope you will use Supervision?
Very cool
@@Roboflow yes, looking into Supervision right now. My computer vision skills need all the help they can get!
One of My fav CV teacher Piotr Skalski
It is Piotr from the video! Thanks a lot!
sooooooooooooo good, please keep developing!!!
What would you like to see next?
it would be really nice if you can display the speed of the car? or the number of cars/people that crossing the road at red /green light@@Roboflow
I personally prefer this format of video where you work from the IDE rather than Colab as I find it easier to follow the structure of your thinking, how you go about solving the problem piece by piece.
Thanks for sharing. Funny enough, you are one of the first people who “voted” for IDE.
Great Video , Please Continue , Thanks Peter :)
Thanks, will do!
Hello Piotr, I hope you are doing fine! I'm at 11:49 min and I'm getting this error:
/python3.9/site-packages/supervision/detection/core.py", line 28, in _validate_mask
raise ValueError("mask must be 3d np.ndarray with (n, H, W) shape")
ValueError: mask must be 3d np.ndarray with (n, H, W) shape
I haven't been able to solve it. I also clone the repo and I got the same problem. 😥
Do you get that with our example? Or your model / your video?
Hi there! I got that from your model and video. I ran the bash script to download both video and model from the drive
Hey@@Roboflow I test the code using windows 11, python 3.11.4 with venv and it did work. I was working with conda in linux.
Yes, Please continue to the project: I like Both IDE and Colab.
Loved this project!
Thanks a lot 🙏🏻
thank you very much for the video and I asked if I wanted to analyze the congestion by determining the length of the traffic jams, the number of cars in traffic jam for each lane of traffic. THANKS!
What about counting traffic on turning movements?
A to B, A to C, A to D and so on
Great video! Next step would be identify the crosswalks and build the polygons with that data
such a great project you've done!
sorry but i couldn't find it, where can i access your code?
A really great tutorial. I have one question though, while running script.py on your video it works fine, but when I run it on my video at 12% it quits with error " in _get_ufunc_and_otypes
raise ValueError('cannot call `vectorize` on size 0 inputs '
ValueError: cannot call `vectorize` on size 0 inputs unless `otypes` is set". Any ideas why?(I added otypes=[object] as second vectorize parameter and I am waiting for 12 hours now.) And how could I display class name alongside tracker id and group the tracking count displayed on zones according to vehicle class?
Why it's not in Google Collab?
You should be able to run that script in Colab
Excellent ! Thanks for sharing your work. Did you finetune a Yolov8x model to detect as well the cars ?
Yes. It was trained on custom dataset.
I am working on a project, for detecting saplings in a field. but when the drone is taking a turn to change the lap (basically moving into other lanes) the previous lanes are also coming like 1,2 of them and saplings in those laps are counted again. Is there any solutions for this?
Did you tried any image stitching? So
I’m short instead of prow easing video. Pick frames every 1 or 2 seconds. Stich them together to form large image. And process that large image?
@@Roboflow actually I have a video from client, I have yolov8 model to detect saplings and the video is shot from drone like if the drone move in first lap, it covers first 4 lanes and when takes a right U turn and move to the 2nd lap the video has previous last 2 or 1 lanes, and saplings on those lanes are counted again, will image stitching work?
hi thanks for the video its awesome just wanna ask how can we use this to detect traffic violations such as wrong way driving
Can you please upload a tutorial on using YOLO-NAS and BYTETrack for object detection and tracking on custom dataset? (if possible, pls include the annotation part for videos of the dataset as well)
Hi we have a separate video covering training for YOLO-NAS. Did you have a chance to watch it? :)
I got a error while compilation with visual code where it state that objects are nun cannot iterate the null objects plz suggest how to solve that issue
can you make videos on this topic Traffic flow tracking at an intersection based on direction-to-direction analysis involves detecting vehicles and determining their movement direction within predefined regions
how is that different from what we showed here?
it only show total no of vehicle not showing classified vehcile number how can we do please ..
From which direction and how many vehicles have gone on signalized intersection with their class name ?" cannot be showed .. how can we place a name in the line and for example D is the one direction A is the another direction how can we show how manh vehicle can pass from d-A
Thank you for this innovative and perfective project! Maybe COLAB version is better option for new starters-low budget developers :)
Everything shown is with open source, free tools! github.com/roboflow/supervision/tree/develop/examples/traffic_analysis + universe.roboflow.com/
I've a " football player tracking and statistics collecting during de play "project to do ... do you've somes ressources for documentation??
Sorry but I’m not sure if I understand correctly. Could you elaborate?
@@RoboflowI want to be able to collect all the statistics during a soccer match, such as the ball procession, the number of corners and so on...
Excellent work and great video, can you please share the model & name of the video camera used?
On GitHub there is .sh script that you can use to download both the model and video.
@@Roboflow I want the name of the video camera used for this testing.
Thank you for a great work! Would be happy to see you continue experimenting more😊
Question though, I assume Polygon instance only contains information whether the ID is in the zone or outside without actually showing whether it crosses the line to exit?
The reason I’m asking this is that my projects has a lot of occlusion, so it’s quite often for model to lose track.
What will happen if object is not tracked anymore, but it never did actually cross the line to exit?
Thank you very much!
So what you are looking for is information that given ID was outside and at some point it cross and is inside? To prevent ID popping out of nowhere in the middle of the zone and get counted in?
Not exactly. I want be sure that actual zone exit (crossing the line) won’t be confused with model losing the object.
Say the object disappeared (due to occlusion) inside the zone. Would the polygon logic treat it as zone exit?
@@Roboflow I'm eager to use and promote your framework. please let me know if you can provide this information:
I want to be sure that actual zone exit (crossing the line) won’t be confused with model losing the object.
Say the object disappeared (due to occlusion) inside the zone. Would the polygon logic treat it as zone exit?
Impressive job!
I have one question, although this video was captured by a drone, the camera lens remains stationary. In most cases, the drone is moving, so fixed-position bounding box is not applicable. Is there a way to address this issue? Thank you!
The script should work with simple camera you just need to put it high enough. It doesn’t need to be drone locate directly over the road.
can you explain how to create the project like the flow from north to east and east to west how can we calculate
Can I run all of this also locally? Or do I really need colab?
I did run everything locally on my MacBook.
what if the video was recorded by a drone cruising back and forth? how to set up the line that keep in a fixed position for flow count?
is the source code shown in this video available in git? if not, kindly provide the source code...it would be of Immense help..TIA
Looks like the source code from the video is here, just found it
github.com/roboflow/supervision/tree/develop/examples/traffic_analysis
I have lost hope in finding the answer of my question @roboflow team. I have been commenting on your last 3-4 videos for one help. I want you to help me with Autodistill auto annotations video,where I want to check the annotated images manually so that i can correct the annotation which is wrong and missing. You told me that you will upload a blog, if its already uploaded please share the link or please make a tutorial on the steps . Thank you .Please i need your help
Hi, I was wondering to know, how to export counting results to csv
We need some projects involving vision transformers, please. One of the applications is desnowing images.
Do you mean detection/segmentation transformers? Or image generative AI?
how much cust ( in hardware components) to implement this code in a custom video (30:00 + duration)?
Hi 👋🏻 I’m not sure I understand what you mean. Could you help me out and explain a bit more?
Do I need to have a very expensive PC to handle the processing that this software consumes? in terms of processor and video card for example@@Roboflow
thank for great tutorial. I want to implement a tracker which counts and track fruits on tree. The problem is some of the fruits are counted more than once. does anybody have any ideas? is there a way to keep the tracklets in ByteTrack for longer duration?
How do I add speed detection in this
Where is jupyter notebook for this project located?
Thank you so much for this video, it is a lot fun learning from your videos!
I have one question, how do I convert the results I have gotten into a output video? I have tried, however I have only gotten the video without any labels or bounding boxes.
oh I GOT IT! Thank you nonetheless for making these awesome videos!
Is it possible for you to create a video whereby ReId is combined with tracking algorithm(eg ByteTrack or BOTsort)?
Thank you once again!
Video from you really helps me a lot I appreciate it . I have a query How can I count number of objects entered and pass from polygon and show the count on polygon
i have this doubt wht shld i enter as --source_weights_path
nice.thx!
My pleasure!
impressive!
Actually it would be nice to draw the zones (bounding box) using mouse dynamically and assign a color, how can that be done? any specific libraries for it?
It possible to do it , I did it modifying Streamlit Drawable Canvas component . and having a Config.json file to feed the code
Can you share it? @harold0723 I’d love to see that!
Please put vedios about adversarial attacks
For "solar panel counting using SAM" ...which process is better >>> 1. detection followed by segmentarion Or 2. segmentation followed by classification
From my experience 1 with YOLOv8 + SAM is suuuuuper easy!
i forgot to mention my images are UAV orthomosaic images...so, is YOLOv8 + SAM still good for it ?
@@Roboflow
Really great project! I would like to save the tracking data for further analysis - speed calculation, trajectory conflict points. What would be the best way to do it? Let's say the result should be a dictionary with all tracked IDs as keys and the values will be a list with the object coordinates in each frame: {id1: [[x1,y1,frame1],[x2,y2,frame2].......} Thank you!
can you help me in making ANPR system where the number plate in detected and there is automatic opening and closing of gate
hey bro,
let me know where I get the code ???????
please
bdw your way of doing code is osm
I trained our network with custom dataset. with YOLOv8 The training period lasted a long time. I want to test the performance of my test set with the network I trained at another time. Is there any other solution than retraining the network?
I’m afraid not. The best you could do is to resume the training. So to pick up where you left of after the first round of training.
Can we run the program with real-time sources on Windows 11 OS? I can't recognize which OS you are working on. Thank you in advance :D
Yes
@@aniketthakur1163did you try it ? I'm getting a lot of errors and I can't figure out please help me if you can
where do I find the code for this ?
Here it is: github.com/roboflow/supervision/tree/develop/examples/traffic_analysis :)
how to know tracklet length?
Can roboflow extract features and label an entireimage as a unique image or can it only label objects within an image
We can do many different things. You can label with single or multiple categories. You can assign tags. You can cluster images based on similarity. Could you be a bit more specific. What solution are you looking for?
@@Roboflow thank you! reverse image matching to find the closest match from an image set using the entire image instead of labeling things inside the image... let me know!
Could anyone let me know what makes the code autotype in the python script? Is it copilot?
Wondering about it too. That was a real speed up over my typical vim coding environment
Yes! I’m using GH Copilot and AI Assistant form JetBrains :)
If you are not using Copilot or any similar plug-in, I highly recommend it.
Hey can you create a tutorial for smart Surveillance which detect human intrusion classify between authorised and unauthorised intruders and notify security person usign YOLO V8 please
Where can I look for source code.
can i get vehicular speed?
Is there any way I can detect speed of the vehicle too and show it on the bounding box itself?
Can anyone help, please?
Sure it is. But you would need to know the real dimensions of the road that the car is traveling on.
I'd love to see how to make a system that tracks people on a, for example, clothing store, and generates a heat map of the most popular spots in the store O.o
So a heat map of the most frequent traveling paths or places where people spent the most time?
@@Roboflow exactly that!! i think both cases would be interesting to see, is it one more difficult than the other?
Nice Video! Thanks for uploading! Can I get the github code for this video?
Hey dude, it's going to sound a bit lame, but my training has failed many times already and I think it's because of my dataset, I need it to be able to detect cars in that same kind of camera angle, so... can you give that database or tell the link where you got it from please 😅
Link to the data set is in the description:)
How to incorporate DeepSort into supervision?
You want to contribute DeepSort into Supervision or implement it in your project?
@@Roboflow implement deepsort in supervision
@Roboflow, all: Has anyone managed to run YOLOv8 on a Coral Edge TPU?
I've converted the YOLOv8 model:
model = YOLO('yolov8n.pt')
model.export(format='edgetpu',int8=True)
I get:
Model successfully compiled but not all operations are supported by the Edge TPU. A percentage of the model will instead run on the CPU, which is slower.
Number of operations that will run on Edge TPU: 230
Number of operations that will run on CPU: 26
Running the model with the "yolov8n_full_integer_quant_edgetpu.tflite" file (and a properly plugged in and recognized Coral Edge TPU USB) results in detection times at least two times slower compared to running on my CPU.
- Are these the correct steps?
- Is this an inherent problem of the YOLOv8 model? i.e. not all of it can be efficiently quantized to run on a Coral Edge TPU?
- What hardware do you guys use for edge real time YOLO?
Sorry I don’t have experience with Coral devices. I have it one of my drawers but newer used it. Maybe we should create YT tutorial about it…
i personally enjoyed the IDE more than Colab and please do the ones you talked about. traffic density, vehicle speed even the gap between vehicles travelling in the same direction. This will be interesting. What if the zones can be entered from the UI instead of hard coding them. just a thought!
Cool project ideas. I’ll keep those in mind. As for zones, with just a little bit of coding you can make them movable or define them from UI. :)
BY THE WAY IS POLYGON ZONE WORKING?
Isn’t it? Did you encounter any problems?
@@Roboflow i never find it when i search...its a broken link
code link?
Yes, IDE is much easier to follow
I’m trying to do both. Some people had different opinion. They prepare colabs because they do not need to setup environment locally.
@@Roboflow Your videos and contributions to the community are incredible, this is just a minor detail. Thank you 🙏🏻
Hi there, we are getting an error of ultralytics. Could you please help me?
Hey Piotr, Huge fan of you but, I have one problem is I'm working in Windows Pycharm. And I have no clue how to run the code in mine ... When i run the code the below part of the code says rebuild it in Windows so if it's possible can you help me in this. Thanks for your work and you keep me motivated to learn more about CV !! Huge fan mate
cv2.imshow("frame", processed_frame)
if cv2.waitKey(1) & 0xFF == ord("q"):
break
cv2.destroyAllWindows()
Hi 👋🏻 thanks a lot! Please ask your question here: github.com/roboflow/supervision It is hard to debug in YT comments.
@@RoboflowI did I hope to get a reply asap
Buen video perrita, tú librería me va a servir mucho. Besos en el aniseto