Видео

Thanks! That sounds really interesting. I might do a different one, I really dove into Pyramids about a year ago, but haven't thought much about them since.

These things can be tricky, the method I'm using here actually tries to "cancel out" the effects of camera motion, so if the drone is stationary then it should work better. If you know that the drone will be stationary for a long time then you can use of these first 3 methods ruclips.net/video/GeqSF8EV1Gs/видео.html

@@pixeloverflow Sorry, I may not have expressed my thoughts clearly. I want to use a camera to detect drones, but the target drone may be stationary, so motion based detection may not be effective. Do you have any ideas to assist in detection?

@@徐志豪-e9u Ok, I think I understand now and that's a tough problem. Here's some quick ideas, let me know if you decide to try any of these. You could run an object detection model (i.e. YOLO), but that would add extra latency. You could use the motion of the drone to determine if something is present. This would be almost the opposite of what I'm doing here, meaning that you would probably reject objects that are moving too much. You could also train your own model to recognize what objects "look like". So every time you detect an object, you mark it as a training sample. This would be a tricky loss function since there are likely many stationary objects that would go undetected and may confuse the model.

I am currently working on small object detection, as motion detection is more favorable for small targets but also brings a lot of interference. However, YOLO's performance on small object detection is less than satisfactory, so I am trying to combine them

@@徐志豪-e9u Yes, I've noticed that it's really challenging to use YOLO for small objects. Have you tried this workflow? blog.roboflow.com/detect-small-objects/

Hi, would you be able to clarify this? I don't remember using any type of license for this.

Glad you enjoyed it!

Sure, send an email to 'pixelflowai@gmail.com' and I'll send it over

No problem, your data can be structured in such that all (x,y,z) positions are accessible in a matrix format (Is this what you meant?). You might need to make some transformations and/or fine tune if your data is much different than the training data.

Thanks!

Hello, I'm glad you enjoyed these! I never got around to formally evaluating the system, from my perspective I was more interested in capturing the 3D location from LiDAR and fusing them with an "off the shelf" YOLOv5 model. In hindsight, not having any metrics wasn't a such good idea.

In the future, I may look into 3D detection on LiDAR point clouds, in this case I will gauge performance with various metrics (I'm not currently sure which ones would be most appropriate here).

Thanks!

Thanks! Glad it helped

Thanks for the feedback, sorry you didn't learn anything, I try to get better with each video. This video is intended to complement this Point Net series that I wrote which assumes you are already familiar with Deep Learning: medium.com/@itberrios6/point-net-from-scratch-78935690e496. Maybe this would help you understand better.

It's possible, I don't have anything planned though

Thanks! Glad you liked it, it took a while for me to figure this out!

I don't have a MATLAB license, but I made this playlist that shows how to fuse Camera and LiDAR in Python. You should be able to convert the concepts into MATLAB. ruclips.net/video/hXpXKRnnM9o/видео.html&pp=iAQB

Let me know if this helps

I had many issues when first setting this up. The way I was able to solve them was to ensure I had the correct dependencies.

If you're still having trouble after following the video, you can post your issue on the Luxonis forum: discuss.luxonis.com/

Hello, I might do that in the future, but in the meantime I have other things planned since most of the modern architectures are available on GitHub and open3d. If your interested in learning more I have an article about KPConv: medium.com/@itberrios6/introducing-kpconv-d0f186172ad5

Glad it was helpful!

It depends on the application, it would be hard to do it for this video, I would say situation where the camera doesn't move is ideal. The tracking algorithm is able to gauge the motion of the objects in terms of pixels, if we know how the pixels correspond to real distances in the scene, then we can compute velocities. We would need to have camera info (camera matrix, calibration parameters, distortion coefficients) to account for lens distortion. A simple analysis is object counting, tracking the objects tells us which ones we have already counted, or else we could end up counting the same object over 100 times in a few seconds.

@@pixeloverflow Thank you for your response and time and effort in making the video :) I just want to point out one thing: the content might need something more than just redoing what a lot of bigger channels already did. Please, don't get me wrong. I highly appreciate your time and effort making the video.

@@nguyenthehoang9148 Thanks! Maybe I'll make a special video with something much different

No problem, thanks for watching!

Thanks for watching!

Thanks!

Thanks!

Thanks! In this video I only used raw KITTI data which has no corresponding "truth" data, but there are several KITTI benchmarks (2D and 3D object detection, tracking, optical flow), with "truth" that can be used. I think an approach to evaluate this work in terms of metrics would be to use the KITTI bench mark data for "Object" and "Depth" to evaluate the estimated depth and the detected objects using this same approach.

@pixeloverflow I understand ground truth data is not available in the kitti raw dataset, but can we apply the same coordinate transformation on Kitti 3D OD data which has labels file in it for evaluation. Also 3D dataset is too big for colab, can you suggest some good alternate for colab. One last thing we can get 3d bounding boxes for this approach since we already have 2D coordinates and depth, right ?

@@manojdhiman574 Yeah that should work. I'm not sure of any good alternates for Colab, I'll try to get back to you if I find anything. For the 3D bounding boxes, yes we can make 3D bounding boxes using the depth data. I've done this before (it's been a while so I don't remember the details), but it is slow and doesn't work very well. I believe Deep Learning approaches currently perform well on this though.

@pixeloverflow Thanks a lot for sharing this approach, I learned something new. Keep up the good work.

@@manojdhiman574 Thanks!

Glad this helped! Thanks for watching!

Thanks for watching! Yes, I'll try to get some C++ implementations in.