First of all, I have looked into the list of all branches of StackExchange, and this seems to be the one best suited for this question.
I am looking for a comparative analysis (can be both theoretical and implementation-oriented) between various popular methods of object recognition and tracking using a Microsoft Kinect 360. These methods do not have to include specialised Kinect API features like hand gesture recognition or skeleton tracking. Could you point me to some technical literature that tackles this subject? I have found quite a few papers that talk about doing object detection and training based on the PCL library. I have also found some papers on using just the RGB image of the Kinect to do classic object tracking. But to put things into perspective, I want to know which method gives good performance and is less challenging to implement when a set of Kinects (possibly with overlapping projection cones) is used to do object recognition and/or tracking.
In my opinion, building a unified point cloud to analyse and label the objects to recognise/classify them using multiple Kinects (assuming I have multiple BUSes) would have too high processing overhead. Would it be a viable alternative to do foreground extraction on the depth images separately and then somehow identify duplicates?
Related
https://learn.microsoft.com/fr-fr/azure/cognitive-services/computer-vision/concept-object-detection
https://cloud.google.com/vision/docs/object-localizer
I would want to know how many and which objects are recognizable using theses APIs and I can't find a mention of that fact.
I found that google API use https://developers.google.com/knowledge-graph/ which is based on schema.org types but I don't really understand well what it's all about.
I'm sorry but as far as I know, there is no fixed list of classes that Azure Computer Vision is able to detect.
By the way, even if there was one, this list is evolving on a regular basis (but no schedule is announced).
In any case, there are limitations (see doc here):
It's important to note the limitations of object detection so you can
avoid or mitigate the effects of false negatives (missed objects) and
limited detail.
Objects are generally not detected if they're small (less than 5% of
the image).
Objects are generally not detected if they're arranged
closely together (a stack of plates, for example).
Objects are not differentiated by brand or product names (different
types of sodas on a store shelf, for example). However, you can get
brand information from an image by using the Brand detection feature.
If you want to detect specific objects, I would highly suggest using Custom Vision (doc / overview here), not Computer Vision, where you can train your model with your own images to match what you are trying to detect
I want to create a application which converts 2d-images/video into a 3d model. While researching on it i found out similar application like Trnio, Scann3D, Qlone,and few others(Though few of them provide poor output 3D model). I also find out about a technology launched by the microsoft research called mobileFusion which showed the same vision i was hoping for my application but these apps were non like that.
Creating a 3D modelling app is complex task, and achieving it to a high standard requires a lot of studying. To point you in the right direction, you most likely want to perform something called Structure-from-Motion(SfM) or Simultaneous Localization and Mapping (SLAM).
If you want to program this yourself OpenCV is a good place to start if you know C++ or Python. A typical pipeline involves; feature extraction and matching, camera pose estimation, triangulation and then optimised using a bundle adjustment. All pipelines for SfM and SLAM follow these general steps (with exceptions of course). All of these steps are possible is OpenCV although Googles Ceres Solver is an excellent open-source bundle adjustment. SfM generally goes onto dense matching which is where you get very dense point clouds which are good for creating meshes. A free open-source pipeline for this is OpenSfM. Another good source for tools is OpenMVG which has all of the tools you need to make a full pipeline.
SLAM is similar to SfM, however, has more of a focus on real-time application and less on absolute accuracy. Applications for this is more centred around robotics where a robot wants to know where it is relative to its environment, but it not so concerned on absolute accuracy. The top SLAM algorithms are ORB-SLAM and LSD-SLAM. Both are open-source and free for you to implement into your own software.
So really it depends what you want... SfM for high accuracy, SLAM for real-time. If you want a good 3D model I would recommend using existing algorithms as they are very good.
The best commercial software in my opinion... Agisoft Photoscan. If you can make anything half as good as this i'd be very impressed. To answer your question what resources will you require. In my opinion, python/c++ skills, the ability to google well and a spare time to read up on photogrammetry and SfM properly.
I am currently working on a (school-)project involving a robot having to navigate a corn field.
We need to make the complete software in NI Labview.
Because of the tasks the robot has to be able to perform the robot has to know it's position.
As sensors we have a 6-DOF IMU, some unrealiable wheel encoders and a 2D laser scanner (SICK TIM351).
Until now I am unable to figure out any algorithms or tutorials, and thus really stuck on this problem.
I am wondering if anyone ever attempted in making SLAM work in labview, and if so are there any examples or explanations to do this?
Or is there perhaps a toolkit for LabVIEW that contains this function/algorithm?
Kind regards,
Jesse Bax
3rd year mechatronic student
As Slavo mentioned, there's the LabVIEW Robotics module that contains algorithms like A* for pathfinding. But there's not very much there that can help you solve the SLAM problem, that I am aware of. The SLAM problem consist of the following parts: Landmark extraction, data association, state estimation and updating of state.
For landmark extraction, you have to pick one or multiple features that you want the robot to recognize. This can for example be a corner or a line(wall in 3D). You can for example use clustering, split and merge or the RANSAC algorithm. I believe your laser scanner extract and store the points in a list sorted by angle, this makes the Split and Merge algorithm very feasible. Although RANSAC is the most accurate of them, but also has a higher complexity. I recommend starting with some optimal data points for testing the line extraction. You can for example put your laser scanner in a small room with straight walls and perform one scan and save it to an array or a file. Make sure the contour is a bit more complex than just four walls. And remove noise either before or after measurement.
I haven't read up on good methods for data association, but you could for example just consider a landmark new if it is a certain distance away from any existing landmarks or update an old landmark if not.
State estimation and updating of state can be achieved with the complementary filter or the Extended Kalman Filter (EKF). EKF is the de facto for nonlinear state estimation [1] and tend to work very well in practice. The theory behind EKF is quite though, but it should be a tad easier to implement. I would recommend using the MathScript module if you are going to program EKF. The point of these two filters are to estimate the position of the robot from the wheel encoders and landmarks extracted from the laser scanner.
As the SLAM problem is a big task, I would recommend program it in multiple smaller SubVI's. So that you can properly test your parts without too much added complexity.
There's also a lot of good papers on SLAM.
http://www.cs.berkeley.edu/~pabbeel/cs287-fa09/readings/Durrant-Whyte_Bailey_SLAM-tutorial-I.pdf
http://ocw.mit.edu/courses/aeronautics-and-astronautics/16-412j-cognitive-robotics-spring-2005/projects/1aslam_blas_repo.pdf
The book "Probabalistic Robotics".
https://wiki.csem.flinders.edu.au/pub/CSEMThesisProjects/ProjectSmit0949/Thesis.pdf
LabVIEW provides LabVIEW Robotics module. There are also plenty of templates for robotics module. Firstly you can check the Starter Kit 2.0 template Which will provide you simple working self driving robot project. You can base on such template and develop your own application from working model, not from scratch.
I want to evaluate the performance of several SDKs / frameworks for depth cameras. These cameras can either be using Time-of-Flight or structured light.
The framework should be capable (at least) of person tracking / blob detection and gesture recognition.
So far I found the following frameworks:
OpenNI (structured light only)
Microsoft Kinect SDK (Kinect only)
Beckon SDK by Omek Interactive (ToF and structured light)
iisu by SoftKinetic (ToF and structured light)
Are there any other frameworks I should be aware of?
EDIT: I found this article by Techradar that seems to indicate that these are indeed the only options currently available.
Any feedback would be very much appreciated!
I have found some interesting links on this. You can take MIT's approach using CodAC . They list lots of facts on this post, the most important ones I will post here.
9. What are limitations of this technique?
The main limitation of our framework is inapplicability to scenes with curvilinear
objects, which would require extensions of the current mathematical model.
Another limitation is that a periodic light source creates a wrap-around error
as it does in other TOF devices. For scenes in which surfaces have high reflectance
or texture variations, availability of a traditional 2D image prior to our data
acquisition allows for improved depth map reconstruction as discussed in our paper.
10. What are advantages of this technique/device and how does it
compare with existing TOF-based range sensing techniques?
In laser scanning, spatial resolution is limited by the scanning time.
TOF cameras do not provide high spatial resolution because they rely on a
low-resolution 2D pixel array of range-sensing pixels. CoDAC is a single-sensor,
high spatial resolution depth camera which works by exploiting the sparsity of natural
scene structure.
11. What is the range resolution and spatial resolution of the CoDAC system?
We have demonstrated sub-centimeter range resolution in our experiments.
This is significantly better than fundamental limit of about 10 cm that would
arise from using a detector with 0.7 nanosecond rise time if we were not using
parametric signal modeling. The improvement in range resolution comes from the
parametric modeling and deconvolution in our framework. We refer the reader to
our publications for complete details and analysis.
We have demonstrated 64-by-64 pixel spatial resolution,
as this is the spatial resolution of our spatial light modulator.
Spatially patterning with a digital micromirror device (DMD) will enable
much higher spatial resolution. Our experiments use only 205 projection patterns,
which correspond to just 5% of number of pixels in the reconstructed depth map.
This is a significant improvement over raster scanning in LIDAR, and it is
obtained without the 2D sensor array used in TOF cameras.
Also another interesting project I found on Youtube uses libfreenect and libusb
There is also dSensingNI which is described as
This work presents an approach to overcome the disadvantages of existing interaction
frameworks and technologies for touch detection and object interaction. The robust and
easy to use framework dSensingNI (Depth Sensing Natural Interaction) is described,
which supports multitouch and tangible interaction with arbitrary objects. It uses
images from a depth-sensing camera and provides tracking of users fingers of palm of
hands and combines this with object interaction, such as grasping, grouping and
stacking, which can be used for advanced interaction techniques.
So you have hit most of them out there, especially that use Kinect, but there are a few other options out there! Hope this Helps!
I'm currently using a Processing Kinect library which supplies a depth map. I was wondering how I could take that and use it to create a 2D skeleton, if possible. Not looking for any code here, just a general process I could use to achieve those results.
Also, given that we've seen this in several of the Kinect games so far, would it be difficult to have multiple skeletons running at once?
Disclaimer: the reason why you still didn't get an answer for this question is probably because that's a current research problem. So I can't give you a direct answer but will try to help with some information and useful resources for this topic.
There are mainly 2 different approaches to create a skeleton from a depth map. The first one is to use machine learning, the second is purely algorithmic.
For the machine learning one, you'd need many samples of people doing a predetermined move, and use those samples to train your favorite learning algorithm. That's the approach that was taken and implemented by Microsoft in the XBox (source), it works really well BUT you need millions of samples to make it reliable... quite a drawback.
The "algorithmic" approach (understand without using a training set) can be done in many different ways and is a research problem. It's often based on modeling the possible body postures and trying to match that with the depth image received. That's the approach that was chosen by PrimeSense (the guys behind the kinect depth camera technology) for their skeleton tracking tool NITE.
The OpenKinect community maintains a wiki where they list some interesting research material about this topic. You might also be interested in this thread on the OpenNI mailing list.
If you're looking for an implementation of a skeleton tracking tool, PrimeSense released NITE (closed source), the one they made: it's part of the OpenNI framework. That's what's used in most of the videos you might have seen that involve skeleton tracking. I think it's able to handle up to 2 skeletons at the same time, but that requires confirmation.
The best solution is to use FAAST (http://projects.ict.usc.edu/mxr/faast/) which requires OpenNI. I have struggled to get OpenNI to work on my computer. I have not seen an approach yet using Code Laboratories' CL NUI.
An algorithmic approach is http://code.google.com/p/skeletonization/ but you may have a problem because your depthmap only represents surfaces and no closed objects.