Another great tutorial. I'm glad you've covered the issues of latency and frames accumulating in the buffer causing crashes, I've been having trouble with that when trying to run CV applications for long periods.
@Roboflow, what was done in the video is great! I loved it!!! . But it's a workaround; you took a video and turned it into an RTSP link. However, there are many challenges that occur when pulling footage from a stream. I would love to see that part done as well. Let's say you get permission for a few cameras from an office/gym or elsewhere and do an example project. It would be fascinating to see the process of pulling footage from these cameras, training a model, thinking of a use case that solves a real-world problem, coding it, and finally giving the result to the client (showing us thats the average time of a customer in checkout 6 is 50 sec on a given day) . Currently, most CV i see is done on 15-second videos and from one camera, and there isn't much value in that. real value is in solving real problems with multiple cameras models/logic that works together. I'm aware that what I'm talking about is a huge project and not an easy task, but if anyone can do it, it's @Roboflow and Piotr. I believe that a tutorial/series like I'm imagining would open the door for millions of CV applications to be built in the future.
Thats exactly what we're doing at the moment for a soccer stadium in UK, taking multiple streams from the CCTV and measuring the average wait time at some turnstiles and food/drink stalls. If the Proof of Concept is successful we will also use CV to monitor occupancy and highlight spare seats in the stadium on a floorplan. Roboflow and Supervision are really useful tools for us!
Today we are hosting a live community session. We will be talking mostly about real-time stream processing and time calculation. Would be awesome if you could join!
Hi, thanks for the really great and easy to understand tutorial. Here, you implemented a clock based timer approach to calculate elapsed time from a real time stream. I am assuming that similar approach can be used to estimate the speed of an object (like car) from a real time stream as well. Is it correct? Would appreciate your valuable feedback.
That’s a very good question. It is guaranteed that you will not drop s because of InferencePipeline logic. You may still drop it because of internet connection. But not because of InferencePipeline. We tested the logic on Jetson Nano decoding 4K stream and it worked really reliably.
Thank you for this amazing tutorial, i just have a problem I'm working with project that detect and tracking people but in my model when people change thier position the model record it's another track so how i can enhance it
I am using yolo8 from ultralytics and bytetrack from supervision. I am doing exactly the same thing, Only difference is I am reading 2 videos at the same time. Yolo detecting humans but tracker doesn't update with that detections. Detection suddenly dissappear. What should be the problem?
@Roboflow Wow this is great content so much exposure to unleash YOLO features. im working on a realtime project which is taking in about 10 stream feeds and has 6 different yolo models(Yolov8) with different usecases. i have currently applied threading on the usecases but im confused how can i feed in 10 streams parallel. im looking for suggestions from on prem deployment point of view as well
Today we are hosting a live community session. It would be so cool if you could join. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-u7XUC-3TqY8.html&ab_channel=Roboflow
Is there a simple way to save class detections in real time to an updating excel file or something similar? i.e. a continuous record of timestamps, with each timestamp containing the object detections, the zone it was detected in and/or bounding box positions/conf at that time, and this information is real-time updating an excel file?
@@Roboflow OK, thank you I thought this is exclusive for the Ultralytics models only, I have already created a custom YOLOv7 & 6 model and I will use them with supervision.
@@Roboflow I have used detect.py, but now I am working on a code to load my model and detect object, in case I want to implement anything with my model it will be easy for me to use my code.
i am not able to use the cuda/gpu support to the code. i tried by using code: model.to('cuda') but getting error. can someone help me on how to implement cuda to the above code. please
is there a way to process high resolution stream in realtime to detect / track small objects in the stream? I have tried SAHI, but its not possible to run in real time if the images are 5000x576.
It will always be a tradeoff between speed and accuracy. You have two options - increase inference resolution(imgsz parameter in YOLOv8) or use SAHI. Both will increase accuracy but decrease speed.
@@Roboflow yes, thats what I also experienced. I am on the way to find a way to do parallel processing of small snippets and then merging the result. Lets see how far I can go
Great video for show, but the processing time is quite slow even if with ultralytics :( Is there any way to speed up the processing time without using GPU ? Thanks in advance!!! You guys are doing a great work in a great direction.
If you don't have a GPU, perhaps you could use a smaller model? Eventually run object detection through OpenVINO - Intel hardware model accelerator. We could talk about it deeper during live community session today. It would be so cool if you could join. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-u7XUC-3TqY8.html&ab_channel=Roboflow
@@Roboflow Sorry for not attending the live session, 10PM is quite late for me cause I am quite tired after a long working day :( ! Hey, I have a proposal like this, can you consult how to achieve this goal (funding, framework, algorithms..., ) docs.google.com/document/d/1ldpQ6q3MpJmox-nOSj2SLq-9sKBMEau1JxHbQClmHMg/edit?usp=sharing . In the past, I used traditional method.ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-N8i2gbh6RRI.html but with the advance of AI, I think we can do further.
@Roboflow, what was done in the video is great! I loved it!!! 😍. But it's a workaround; you took a video and turned it into an RTSP link. However, there are many challenges that occur when pulling footage from a stream. I would love to see that part done as well. Let's say you get permission for a few cameras from an office/gym or elsewhere and do an example project. It would be fascinating to see the process of pulling footage from these cameras, training a model, thinking of a use case that solves a real-world problem, coding it, and finally giving the result to the client (showing us thats the average time of a customer in checkout 6 is 50 sec on a given day) . Currently, most CV i see is done on 15-second videos and from one camera, and there isn't much value in that. real value is in solving real problems with multiple cameras models/logic that works together. I'm aware that what I'm talking about is a huge project and not an easy task, but if anyone can do it, it's @Roboflow and Piotr. I believe that a tutorial/series like I'm imagining would open the door for millions of CV applications to be built in the future.
@@ItayHilel there is simple no data like this that I can use in tutorials. I'd need footage from multiple cameras from store or gym. And people do not share that online :/