YOLOv4 in the CLOUD: Build Object Tracking Using DeepSORT in Google Colab (FREE GPU)
HTML-код
- Опубликовано: 15 сен 2024
- Learn how to build and run your very own Object Tracker in Google Colab! This tutorial walks you through the process of building an object tracking application using DeepSORT and YOLOv4 Object Detection in the Cloud on Google Colab's FREE GPU! Get it up and running with only a few clicks of a button. This tutorial covers it all.
#yolov4 #deepsort #cloud
THE GOOGLE COLAB NOTEBOOK: colab.research...
YOLOv4 is a state of the art algorithm that uses deep convolutional neural networks to perform object detections. We can take the output of YOLOv4 feed these object detections into Deep SORT (Simple Online and Realtime Tracking with a Deep Association Metric) in order to create a highly accurate object tracker.
GET THE CODE HERE: github.com/the...
In this video I cover:
1. Setting up the Colab Notebook and Enabling GPU
2. Cloning the code and installing dependencies.
3. Converting YOLOv4 pre-trained model into TensorFlow model.
4. Running Object Tracker on video.
5. Filtering allowed classes to track.
6. Adding info flag to see detailed information on tracked objects.
------------------------------Resources------------------------------
Run DeepSORT Object Tracker on Local Machine: • Object Tracking Using ...
Learn to Convert to TFLite and TensorRT: • YOLOv4 Object Detectio...
Configure to Run with Custom YOLOv4 Detector: • How to Build a Custom ...
Train Custom YOLOv4 Detector in Cloud: • YOLOv4 in the CLOUD: B...
The Official YOLOv4 paper: arxiv.org/abs/...
If you enjoyed the video, toss it a like! 👍
To Subscribe: / @theaiguy
Thanks so much for watching!
- The AI Guy
Great video and tutorial! Just an FYI: I noticed today that my Google Colab had tensorflow version 2.4.0 with which the converted weights file was smaller (9 MB instead of 11MB) and the object detection + tracking were not working i.e. not found any persons, cars, etc. Then I reinstalled tensorflow version 2.3.0-rc and with that it was working properly. So if anyone were to run into the same problem check your tensorflow version in Colab.
thanks that worked! keep contributing :)
Great tip, this fixed it for me as well, Thank you!
Here is the code that worked for me to backdate tensorflow:
!pip uninstall tensorflow -y
!pip install tensorflow==2.3.0
import tensorflow
tensorflow.__version__
@@ThePlymouth37 Thankyou, now its working
@@ThePlymouth37 Thanks for the tips. Its working now.
Hope you enjoy! Let me know if you like the Google Colab videos or if you would like to see more of them? :)
Very cool, could you show how to use multi/cameras ? I want to track persons with differents cameras
I enjoy this colab videos. I am interested in custom object re-identification integrated with YOLO. Can you make a video on that?
thank you so much! realy enjoy your videos! instance segmentation from annotating data to training custom model maybe a possible one?
Colab tutorials are really useful to quickly play around with the code quickly. What would also be useful would be more documentation explaining what are some of the thresholds/metrics/flags we can tune in Tracker.py/object_tracker.py according to our needs.
For example, I am trying to modify this code to count the number of people who have walked past a certain checkpoint within the video, so I don't necessarily need the tracking boxes to be always present (I guess some false negatives are ok) but I really don't want ID switching which would cause double counting. Would be helpful to know which are the thresholds/metrics I should be tuning.
pls make video on crowd analytics, counting people in a crowd and uploading results on cloud...
Just wanted to comment in case anyone else has this issue, while attempting to use my own custom weights DeepSort would not detect anything. I had to uncomment the dependency line, restart the kernel, re-convert the weights again and then it worked! Also wanted to say that I love this channel, thanks for everything!
i have the same issue, could you tell where to uncomment to depedency line? and how to re-convert the weights?
@@rajkumarayyalsamy1971 uncomment the "!pip install requiremts " line, and once you restart the session you should execute the save_model.py again to convert yolov4 darknet wights into tensorflow model.
thank you so much
Great video and very good channel, I'm a final semester student working on a project using YOLO and this channel is very helpful.
finally, the video I was waiting for, thank you for understanding me, going to follow carefully step by step, thank you again.
Hope you enjoy! Thanks for the support.
Extremely interesting videos, would love for you to tackle multi-label image classification using google colab next. Thankyou for these tutorials.
Make a Discord server, so everyone can connect with you and share some naive or good doubts which can be solved with collective intelligence.
Thanks for putting so much efforts in making great content as always.
I will definitely look into setting this up. Thanks for the great suggestion.
@@TheAIGuy that would be awesome.👍
Thanks a lot! you have an amazing channel. Keep on the good work!
Hey TheAIGuy, you're awesome!!! You're saving in my job so hard haha. Thank you so much
Thank you very much!!!
Is it possible to add custom classes to this object tracker just like you did with the object detector???
Awesome video. I had to uninstall Tensorflow version 2..0 and installed version 2.3.0 after suffering for a week
Hey can you make a video on configuring deepsort to train it on a custom data set ?. Love your content .
this better than my computer class which costs $500
Thank you so much for your video! So easy to follow!
For those having issues, pls try the below
! pip uninstall tensorflow
! pip install tensorflow==2.3.0
amazing video! please make a video on segmentation like segnet and mask RCNN
Great suggestion! Definitely will be covering this in the future so stay tuned.
Para crear la red en base a tu propio dataset: Descarga tus 'pesos' obtenidos en el entrenamiento de YOLOv4 en este nuevo entorno, pon este archivo en la carpeta /data, después deberás crear en esa misma carpeta otro archivo '.names', en mi caso "obj.names", con los nombres de las clases, justo como los pusiste al entrenar YOLOv4. Por último deberás ir a la carpeta /core al archivo 'config.py' y deberás cambiar la linea "__C.YOLO.CLASSES" para que ahora apunte al archivo con los nombres de tu dataset: = "./data/obj.names" en mi caso.
My video has no trackers.
My log is not showing any trackers either
Frame #: 1
FPS: 0.03
Frame #: 2
FPS: 24.59
Frame #: 3
FPS: 27.17
Frame #: 4
FPS: 27.05
Frame #: 5
FPS: 27.36
Frame #: 6
FPS: 27.42
Frame #: 7
Don't see any errors.
Can you please help!
same problem 😪😪
Thanks man! Excellent video, keep it up
great explanation with a good understanding thanks!!!
Thanks @The AI Guy for your nice work on yolo and deep sort! I really enjoyed the videos you did on the subjects. Best youtube content on the topic!
Quick question: I can see that if I try to limit FPS in my yolo+deepsort pipeline, I limit the capacity of deep sort to guess similarities between bounding boxes in order to group them under the same id. Do you have an idea in order to improve the quality of this part of the pipeline? I tried to improve the FPS gradually to guess kind of the limit for now. Is there a temporal parameter in order for deep sort to perform better with longer wholes between frames, or is it just a question of training?
Coooollllllllll.
Hi! Great Video! Can i ask if it is able to use the weights/model from custom detect (colab) and use the weight for deepsort (colab)? If so how can i do so.
Hi Lim, did you get a solution to this problem?
Cool stuff, just one intriguing thing is that, if deep-sort is trained in MARS with only humans, how the model worked really well with cars??
Great Work.....
One Query.... How to integrate Faster RCNN object detection with deep SORT
Thanks for the tutorial! It was quite helpful. But could you make a video on how to build Object Tracking using DeepSORT?
I mean I have followed your video "YOLOv4 in the CLOUD: Build and Train Custom Object Detector (FREE GPU)" and "How to Build a Custom YOLOv4 Object Detector using TensorFlow (License Plate Detector)". They work just fine. But when I try to deploy that self-trainied yolov4-obj_last.weights with DeepSORT, it just didn't work.
Have no idea what's going wrong about it. Thanks anyway!
@TheAIGuy
Thanks for the video, how can we do standard detection on colab using Yolo for Live video
if you are using this repo, you just need to change your input video path as "0"
Hello Sir, I have learnt use of YOLO by watching your videos. Thank you for providing such videos. Sir can you please provide us video on how to distinguish between static and moving object using YOLO in google colab?
Why did I follow every step of your video, and the final output video didn't produce a correctly tracked bounding box like you did?
Hi AI Guy! Good job!!! One question: Where are saved the output BBox coordinates, or how to extract them and write to excel sheet? If anyone can help please do me that favor. Thanks!
pls make video on crowd analytics, counting people in a crowd and uploading results on cloud...
Sounds like a great idea! Thanks for the suggestion!
this will be amazing
Great video with nice explanation
Question: How to apply YOLO in AirSim for object Detection and Tracking
Great work @TheAIGuy, please someone help, My tracker does not show label names, neither Ids, What i happening? Do I need to bring the detector model into the tracking model?
Great video Please do a video on object detection using Densenet
Great video!!!!!!!!!!!
Great work!!!!!!!!!!!.
Just a quick question if anyone is also working on it. How do I get the coordinates of the bounding box getting tracked for the last car video he just ran.
colab is not working for me, it is just making video copy, without any bounding boxes around objects.
Thanks for sharing this wonderful video. I was wondering what is the resolution of the videos that were in your experiment, and how you computed the fps as both the yolo detector and DeepSort tracker also need to take time to unpacked the bounding boxes objects particularly when the numbers are relatively high.
What about doing this in real time with IP webcams ?
Hi I'm running this implementation on Google Colab and getting only 0.5fps processing speed. It would seem the GPU is not being used even though it's available. Tensorflow 2.3.0cr0 has been discontinued but it's still listed in your requirements.txt. Is there an update to get this working? Thanks
Please make a video on semantic segmentation or on Detectron 2.
another fabulous viedo from AI guy
btw, could you do some video about Mask RCNN, it should be fun and helpful for many people
I got error 'Video has ended or failed, try a different video format!' please help!
Thank you. My question is we need to train yolov4 and deepsort separately, right (if we want to fine tune both of them for our application)?
Can calab listen for UDP video stream and return JSON objects for each frame being processed??? On a callback URL??
UDP to my local machine - forward UDP video traffic to calab -calab makes url callback with JSON object of what was detected.
Great job, I am thinking about the motorbike thief detection project. Would you please advise me between 2 options: first one is use posenet and yolov4 then base on the wrists position to detect the malicious chance (this part is coding condition). Second option uses RNN to detect the series of action through time and predict the thief (this one is 100% AI but data is my concern).
Am failing to execute same code on colab, it requires to change the video format. After running the code it does not detect
I tried the same techniques on google collab to make it work (cloning, pip install with your requirements file, placing the weight file in the correct folder ect...) before you posted the video, but it wasn't working when i was trying to save the model (an operation failed because of a problem of dimensions). Now it works perfectly but I am wondering why it couldn't work before, do you know what had to be done ?
Earth Shaker no i didnt change any of the commands. That is strange!
Anyone experiencing an issue with th object tracker not working. It worked before but recently the object tracker doesnt want to run
I watched your video about how to build and train our custom object detector and it was very helpful but i have a question
I used google to train the model then i downloaded the weights and used opencv function dnn_detectionmodel it was good at detection but it's slow and i tried the same thing using frozen inference graph .pb(pre-trained) as weights path and it was fast and i tried to have .pb file from my yolo weights (using save_model.py)then i tried doing the same (using cv2.dnn_detectionmodel ) but it doesn't work
So the question is can we extract frozen inference graph.pb from our yolo weights and use them with opencv to have best results
Thanks in advance
Hello Thank you for these amazing videos.
Could you please make a video about communication between 2 agents, detecting objects?
Hi . Appreciate your videos. Well done. Could you answer one thing ? in one of the videos you have shown the use of OIDv4_ToolKit to download labeled pictures . However I didn't get how to download all labels for one picture. For example I'm doing detector of cars and people, and i want both labels if they exist on one image, not separately labeled two categories of images . Is there the way to download them?
My GPU Memory doesn’t seems to go high and stuck with around 5 to 7 FPS. I ran real time gpu memory tracking code beside and it shows only 6GB of gpu memory is used by this 12GB GPU. Any advices?
Thank you so much for the video and code! Unfortunatelly i've had some problem with the requirements that stopped the making of the video: the tensorflow-gpu requirements and pip v19.0 are in conflict with other versions or settings in my colab. How can i solve this?
Hey there! were you able to solve this issue, I'm having the same problem :(
update the gpu-requirements.txt tofollowing:
tensorflow-gpu==2.3.0rc0
opencv-python==4.1.1.26
lxml
tqdm
absl-py
matplotlib
easydict
pillow
numpy==1.19.2
tensorboard==2.4
folium==0.2.1
imgaug==0.2.7
bro could you explain why we need to do that?
Hi there, first of all thank you for this tutorial. Is it possible to change the code so I can select ROI on the first frame of the video manually and then let the tracker do its job? Or what if I only need to track few people in the video and not the rest of them?
Thanks for the awesome tutorial
Can we add custom ROI lines for counting the number of cars for upstream and downstream traffic
Great video! Thanks for sharing it! It is very clear and precise.
I wanted to ask how do you extract statistics about the object being tracked such as distance the object has covered, velocity. Is this part of deepSORT result that can be specified in the return call from the function? Also how does one add custom classes and bounding box for a new custom class?
in 100th iteration of object detection is gives error and don't save the weights in Drive
Hi TheAIGuy, I enjoyed your tutorials very much.
Would appreciate it if you could do a tutorial on converting darknet yolov4 to tensorRT and run the obj detection with deepsort on jetson nano/xavier nx. Many thanks in advance
Hi please if i want to run this on a different video how will I go about it. I tried inputting the video but it didn't work
Which one is better for real time detection? Using YOLO per frame or using tracking algorithm like deepSORT? I got 1 FPS while rendering the output in my local machine with DeepSORT (GT 755M 2GB GDDR5 GPU)....But my end application will be on webcam video and on my local machine it gave me like 1-2 FPS on webcam lol...Hope it will be better using colab
what a wonderful tutorial..!! i really appreciate it, Thank you...!! may i ask one thing? i wonder that RealTime Tracking Possible with DeepSort? i tried to set to 0 for webcam but not working in colab.
Pls make a video on Crowd counting... 🙏
How can I display the total number of people on the frame?
Thank you for sharing this high level content videos. On Step5 output i get this WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model..... And on STEP6 i didn´t get any detection. Any one with same problem ?
Great video I want to learn more about yolo can yu tell me some resource
Ashutosh Shukla a great first resource is the official YOLOv4 research paper! You can find a link to it in the description of this video. It covers a lot of cool information.
Can you help out, how to estimate the speed of the cars which was in Video.
I love your videos, they are really easy to follow along with. I'm kinda teaching myself about robotics and computers, and I still have a lot to learn. I'm trying to build a home assistant robot and I have a couple microcontrollers (esp32 cam, and an Arduino Uno) I was hoping there was a way that I could stream video to google colab and have it do the processing and then send the results back to the (robot) to execute a command... Is it possible to do this some how? or is the robot going to need its own GPU for this??
Hello, how can i save the output ids and positions in a JSON file?
Hi Ai guy..Thanks for making such a mind-blowing video..I was just curious if we can tracks bubbles captured using high speed camera. Is there any class available for bubble detection?
Amazingggggg, u are my idol
What is the benefit if I buy a GPU? Looks like I can code on Colab.
I followed the instructions but yolo is not working. can you check it again?? Thanks.
How can I train with my dataset (tracking fish)
You will need to train your own custom Yolov4 model using your dataset. The tracker can then be configured to run with your custom model.
@@TheAIGuy it will be great if you make a video about tracking using custom dataset
@@athirapmohan6496 it is very easy to do. You can find the detailed explanation in the code’s Readme.
Hey did you figure this out? Is it necessary to train deep sort for the custom dataset as well. I track falling rocks..
@@giorgossartzetakis8771 no. Just need the weight file of yolo object detection trained with our custom dataset
hai thanks to tutrial yolov4
i have i question can this running in webcam video in realtime?
I have a collab request for soccer shot detection on custom datasets, been having trouble -- any way to contact you ?
Hello sir,
When I follow your code in the colab, the model can't predict. That means the output video looks like an input video without a bounding box with a label.
Thank you for your reply.
me too ，did you solve the problem?
hi... Is it possible to combine two custom dataset
I am fan...can you elaborate on the join.. what comes with each package.. if there is a schedule so we can use it to its fullest if it fix my schedule... we would love to support this channel
Overall, the join is just a way to show your support of the channel. The basic tier is just people who want to support the channel and encourage me to make content. The 'Big Fan' tier is for those who get a lot out of the channel. I will be posting photos of projects as well as project updates for that tier.
@@TheAIGuy Will do that
@@michaelmutekeri8770 Much appreciated! Means a lot, thanks :)
how it works with webcam, does not work with webcam
Hi. hi can I implement this but with social distance tracking and facemask detection? I hope someone can help me out.
Hey is it possible to run this with tensorRT in Jetson Nano?
Thanks for the video! I got this working but it's using CPU and only getting 2 FPS 😳
I have a GPU but don't know how to activate it. Can you help?
Thanks!!
Please make a video how to add new classes in this yolov4 DeepSORT model
Hi, First of all you provide a very good demonstration how it works and all the codes.
Currently I am using your colab notebook and all the videos of it, here the detection is not working on the first step. Nothing is detected.
I haven't done any modification in the notebook
Please check for the same.
Hi Diptesh, did you get a solution to this problem? I need the same assistance
can you pls explain how cloud is used in this video ?
Google Colab notebook’s use a free Cloud VM with a GPU. So when you run the code it is running on one of Google’s machines and not your local one. Hope this helps!
Hi, I really enjoy with your works. Thanks for these works. I have stucked with some multiple object tracking problems. I don't know how to prepare a train and test datasets for evaluation of multiple object tracking process (for example, I am using deepsort). Do you know how can I prepare a test set in order to evaluate yolov4+deepsort?
@The AI Guy hi :)
how tu use in yolov5 ?
Thanks for the detailed video! However, for some reason no classes are detected on step 6. I can see the frame and fps which means that the video is indeed processed, but at the end it says 'Video has ended or failed, try a different video format!' and nothing has been detected. Could you please help? Thanks in advance!
Hi Marita. I got the same problem. Did you manage to solve?
@@eliastemponi6139Hi Elias! I actually didn’t, but I used his other video on yolov4 in the cloud (YOLOv4 in the CLOUD: Install and Run Object Detector) and it worked perfectly! It does not use DeepSORT or anything like that, but it works for me. Also in the next video of the one I mentioned, he describes in detail how to train on a custom dataset. Good luck on your project!
I too have same problem, did u get solution up to yet?
its giving an opencv error when i do it on my own videos !
hye the AI Guy, can you help me to solve the code count person only?
WARNING:tensorflow:Compiled the loaded model, but the compiled metrics have yet to be built. model.compile_metrics will be empty until you train or evaluate the model. this warning solution
! pip uninstall tensorflow
! pip install tensorflow==2.3.0
Can anyone help me with how to add up and down function to this.
Currently the code creates a video that doesnt track any object !
Hello, I have been executing the cells without making any modifications but it does not detect anything at all, nor does it show any error and the same happens to me in the yolo v4 crop video
did you get solution?I got similar problem
@@lilianleonidas9565 my problem was that my custom weights were poorly trained, I tried a test with the originals and it worked and after training my own weights properly it also worked.
@@AndreF0H am still failing to execute same code, as it tells me i should change the video format. can you help how you ran the code?
@@AndreF0H Also for custom dataset which place needs changes apart from weights folder?
cool! but could we run object tracking on the webcam live on colab using yolov4?
when I build my own deepsort tracker on colab. it always says"UnboundLocalError: local variable 'ANCHORS' referenced before assignment"
can do in image
What gpu?
Is it possible to do object tracking live? If so, can you teach us?
use the streams from rtsp cameras