I have a custom USB camera with a custom driver on a custom board Nvidia Jetson TX2 that is not detected through openpose examples. I access the data using GStreamer custom source. I currently pull frames into a CV mat, color convert them and feed into OpenPose on a per picture basis, it works fine but 30 - 40% slower than a comparable video stream from a plug and play camera. I would like to explore things like tracking that is available for streams since Im trying to maximize the fps. I believe the stream feed is superior due to better (continuous) use of the GPU.
In particular the speedup would come at confidence expense and would be addressed later. 1 frame goes through pose estimation and 3 - 4 subsequent frames are just tracking the object with decreasing confidence levels. I tried that on a plug and play camera and openpose example and the results were somewhat satisfactory.
The point where I stumbled is that I can put the video stream into CV VideoCapture but I do not know, however, how to provide the CV video capture to OpenPose for processing.
If there is a better way to do it, I am happy to try different things but the bottom line is that the custom camera stays (I know ;/). Solutions to the issue described or different ideas are welcome.
Things I already tried:
Lower resolution of the camera (the camera crops below certain res instead of binning so cant really go below 1920x1080, its a 40+ MegaPixel video camera by the way)
use CUDA to shrink the image before feeding it to OpenPose (the shrink + pose estimation time was virtually equivalent to the pose estimation on the original image)
since the camera view is static, check for changes between frames, crop the image down to the area that changed and run pose estimation on that section (10% speedup, high risk of missing something)
Related
How to recognize rain on camera vision using with OpenCV in C++?
Or if somebody stick a sticker on a camera how recognize it with OpenCV in C++?
Or if somebody throw color to the camera how can i detect it with OpenCV in C++?
Detect these on camera vision:
Rain
Sticker
Color
Here is an example video of sticker!
Camera Vision-Sticker
In case of a sticker, you're just looking for a large dark area that doesn't change in time.
In case of color, analyze image color stats - if somebody sprays some paint on a camera (is that what you mean by "throwing color"?), some color is going to be dominant over all the others.
You can also try to handle both cases by subtracting frames and detecting image areas that don't change in time that way.
You may want to use machine learning for finding threshold values (e.g. area size, its shape properties, such as width/length ratio, continuousness etc.) used to decide when to consider something to be a sticker/color or something else.
As for the rain, I guess there's no simple answer that can be given in a few sentences. There are some articles available in the web though. That said, I would guess it would be simpler and cheaper to detect rain by just installing external rain sensors (like the ones activating wipers in a car) rather than trying to do it by developing your own computer vision algorithm for that purpose.
This sounds like an interesting project, where a camera can automatically detect obstruction (paint, sticker, rain). It will most likely be necessary for the camera to be mounted without obstructions so that the expected image can be learned. If the usage scenario allows that, it won't be very hard.Both sticker and rain result in strong permanent deviations from the expected image, while rain will result in noisy images.
OpenCV with C++ or Python can help solve this kind of problems, because complicated computer vision algorithms are already implemented there. It takes some time to get started with, but after that OpenCV is not hard.
I want to build a depth camera that finds out any image from particular distance. I have already read the following link.
http://www.i-programmer.info/news/194-kinect/7641-microsoft-research-shows-how-to-turn-any-camera-into-a-depth-camera.html
https://jahya.net/blog/how-depth-sensor-works-in-5-minutes/
But couldn't understand clearly which hardware requirements need & how to integrated into all together?
Thanks
Certainly, a depth sensor needs an IR sensor, just like in Kinect or Asus Xtion and other cameras available that provides the depth or range image. However, Microsoft came up with machine learning techniques and using algorithmic modification and research which you can find here. Also here is a video link which shows the mobile camera that has been modified to get depth rendering. But some hardware changes might be necessary if you make a standalone 2D camera into a new performing device. So I would suggest you to see the hardware design of the existing market devices as well.
one way or the other you would need two angles to the same points to get a depth. So search for depth sensors and examples e.g. kinect with ros or openCV or here
also you could transfere two camera streams into a point cloud but that's another story
Here's what I know:
3D Cameras
RGBD and Stereoscopic cameras are popular for these applications but are not always practical / available. I've prototyped with Kinects (v1,v2) and intel cameras (r200,d435). Certainly those are preferred even today.
2D Cameras
IF YOU WANT TO USE RGB DATA FOR DEPTH INFO then you need to have an algorithm that will process the math for each frame; try an RGB SLAM. A good algo will not process ALL the data every frame but it will process all the data once and then look for clues to support evidence of changes to your scene. A number of BIG companies have already done this (it's not that difficult if you have a big team w big money) think Google, Apple, MSFT, etc etc.
Good luck out there, make something amazing!
I know the title is a bit vague but I'm not sure how else to describe it.
CentOS with ffmpeg + OpenCV 2.4.9. I'm working on a simple motion detection system which uses a stream from an IP camera (h264).
Once in a while the stream hiccups and throws in a "bad frame" (see pic-bad.png link below). The problem is, these frames vary largely from the previous frames and causes a "motion" event to get triggered even though no actual motion occured.
The pictures below will explain the problem.
Good frame (motion captured):
Bad frame (no motion, just a broken frame):
The bad frame gets caught randomly. I guess I can make a bad frame detector by analyzing (looping) through the pixels going down from a certain position to see if they are all the same, but I'm wondering if there is any other, more efficient, "by the book" approach to detecting these types of bad frames and just skipping over them.
Thank You!
EDIT UPDATE:
The frame is grabbed using a C++ motion detection program via cvQueryFrame(camera); so I do not directly interface with ffmpeg, OpenCV does it on the backend. I'm using the latest version of ffmpeg compiled from git source. All of the libraries are also up to date (h264, etc, all downloaded and compiled yesterday). The data is coming from an RTSP stream (ffserver). I've tested over multiple cameras (dahua 1 - 3 MP models) and the frame glitch is pretty persistent across all of them, although it doesn't happen continuously, just once on a while (ex: once every 10 minutes).
What comes to my mind in first approach is to check dissimilarity between example of valid frame and the one we are checking by counting the pixels that are not the same. Dividing this number by the area we get percentage which measures dissimilarity. I would guess above 0.5 we can say that tested frame is invalid because it differs too much from the example of valid one.
This assumption is only appropriate if you have a static camera (it does not move) and the objects which can move in front of it are not in the shortest distance (depends from focal length, but if you have e.g. wide lenses so objects should not appear less than 30 cm in front of camera to prevent situation that objects "jumps" into a frame from nowhere and has it size bigger that 50% of frame area).
Here you have opencv function which does what I said. In fact you can adjust dissimilarity coefficient more large if you think motion changes will be more rapid. Please notice that first parameter should be an example of valid frame.
bool IsBadFrame(const cv::Mat &goodFrame, const cv::Mat &nextFrame) {
// assert(goodFrame.size() == nextFrame.size())
cv::Mat g, g2;
cv::cvtColor(goodFrame, g, CV_BGR2GRAY);
cv::cvtColor(nextFrame, g2, CV_BGR2GRAY);
cv::Mat diff = g2 != g;
float similarity = (float)cv::countNonZero(diff) / (goodFrame.size().height * goodFrame.size().width);
return similarity > 0.5f;
}
You do not mention if you use ffmpeg command line or libraries, but in the latter case you can check the bad frame flag (I forgot its exact description) and simply ignore those frames.
remove waitKey(50) or change it to waitKey(1). I think opencv does not spawn a new thread to perform capture. so when there is a pause, it confuses the buffer management routines, causing bad frames..maybe?
I have dahua cameras and observed that with higher delay, bad frames are observed. And they go away completely with waitKey(1). The pause does not necessarily need to come from waitKey. Calling routines also cause such pauses and result in bad frames if they are taking long enough.
This means that there should be minimum pause between consecutive frame grabs.the solution would be to use two threads to perform capture and processing separately.
I am working on a project that requires me to detect and track a human in a live video from a webcam connected to a Beagleboard xm.
I have completed this task using Opencv in pixel domain. The results on the board are very accurate but extremely slow. Many people have suggested me to leave pixel domain and do the same task in an h.264/MPEG-4 compressed video as it would extremely reduce the computational overhead.
I have read many research papers but failed to discover any software platform or a library that I can use to analyze and process h.264 compressed videos.
I will be thankful if someone can suggest me some library for h.264 compressed video analysis and guide me further.
Thanks and Regards.
I'm not sure how practical this really is (I've never tried to do it), but my guess would be that what they're referring to would be looking for a block of macro-blocks that all have (nearly) identical motion vectors.
For example, let's assume you have a camera that's not panning, and the picture shows a car driving across the screen. Looking at the motion vectors, you should have a (roughly) car-shaped bunch of macro-blocks that all have similar motion vectors (denoting the motion of the car). Then, rather than look at the entire picture for your object of interest, you can look at that block in isolation and try to identify it. Likewise, if the camera was panning with the car, you'd have a car-shaped block with small motion vectors, and most of the background would have similar motion vectors in the opposite direction of the car's movement.
Note, however, that this is likely to be imprecise at best. Just for example, let's assume our mythical car as driving in front of a brick building, with its headlights illuminating some of the bricks. In this case, a brick in one picture might (easily) not point back at the same brick in the previous picture, but instead point at the brick in the previous picture that happened to be illuminated about the same. The bricks are enough alike that the closest match will depend more on illumination than the brick itself.
You may be able, eventually, to parse and determine that h.264 has an object, but this will not be "object tracking" like your looking for. openCV is excellent software and what it does best. Have you considered scaling the video down to a smaller resolution for easier analysis by openCV?
I think you are highly over estimating the computing power of this $45 computer. Object recognition and tracking is VERY hard computationally speaking. I would start by seeing how many frames per second your board can track and optimize from there. Start looking at where your bottlenecks are, you may be better off processing raw video instead of having to decode h.264 video first. Again, RAW video takes a LOT of RAM, and processing through that takes a LOT of CPU.
Minimize overhead from decoding video, minimize RAM overhead by scaling down the video before analysis, but in the end, your asking a LOT from a 1ghz, 32bit ARM processor.
FFMPEG is a very old library that is not being supported now a days. It has very limited capabilities in terms of processing and object tracking in h.264 compressed video. Most of the commands usually are outdated.
The best thing would be to study h.264 thoroughly and then try to implement your own API in some language like Java or c#.
Right now, what I'm trying to do is to make a new GUI, essentially a software using directX (more exact, direct3D), that display streaming images from Axis IP cameras.
For the time being I figured that the flow for the entire program would be like this:
1. Get the Axis program to get streaming images
2. Pass the images to the Direct3D program.
3. Display the program, on the screen.
Currently I have made a somewhat basic Direct3D app that loads and display video frames from avi videos(for testing). I dunno how to load images directly from videos using DirectX, so I used OpenCV to save frames from the video and have DX upload them up. Very slow.
Right now I have some unclear things:
1. How to Get an Axis program that works in C++ (gonna look up examples later, prolly no big deal)
2. How to upload images directly from the Axis IP camera program.
So guys, do you have any recommendations or suggestions on how to make my program work more efficiently? Anything just let me know.
Well you may find it faster to use directshow and add a custom renderer at the far end that, directly, copies the decompressed video data directly to a Direct3D texture.
Its well worth double buffering that texture. ie have texture 0 displaying and texture 1 being uploaded too and then swap the 2 over when a new frame is available (ie display texture 1 while uploading to texture 0).
This way you can de-couple the video frame rate from the rendering frame rate which makes dropped frames a little easier to handle.
I use in-place update of Direct3D textures (using IDirect3DTexture9::LockRect) and it works very fast. What part of your program works slow?
For capture images from Axis cams you may use iPSi c++ library: http://sourceforge.net/projects/ipsi/
It can be used for capturing images and control camera zoom and rotation (if available).