Requirement Analysis for DemoLF

Posted on by
Reply

I want to design a demo program on finding the landing sites for an UAV, which I call it the DemoLF.  I want this program has following features:

(1) Open a video file as a data source for demo program, together with a log file, in which the timestamp of each video frame, GPS data, altitude information, etc has been recorded;

(2) Can show the video on a independant window;

(3) I can write my algorithms into plugins, which can then be loaded and executed by the DemoLF;

(4) The program allow the interaction between software and users, through key binding mechanism. That is: the user hit a key (or key combination) when the program runs, the program will do something as a response, and the key binding can be changed through a configuration file;

(5) Allow each plugin has its own windows for showing the image data / result.

(6) Has debug functions, that is: show some debuging information, and maybe some other functionalities, not for sure;

(7) Configurable.

美国常用网址

Posted on by
Reply

购买机票：

1.       http://www.cheapoair.com

2.       http://www.kayak.com/buzz

3.       http://www.flychina.com/

4.       http://www.orbitz.com  （也便宜）

5.       http://www.priceline.com（便宜）

6.       http://www.expedia.com/

7.       http://www.fly.com

8.       http://www.dealbase.com/

9.       http://www.bookingbuddy.com/(含多个网站，可以比较)

10.   https://www.onetravel.com/

看电影：

http://www.cinemark.com/

租车：

http://www.enterprise.com/

旅游：

http://cn.toursforfun.com/

政府网站

http://www.cstx.gov/

http://www.bryantx.gov/

 旧货（车等）网站

http://www.craigslist.org

订shuttle：

www.groundshuttle.com

 积分网站，买机票

www.huaren.us

Seaworld tickets

http://www.ttia.org/?page=tickets

 

Note: Reinforcement learning algorithms for robotic navigation in dynamic environments

Posted on by
Reply

[1] G. G. Yen and T. W. Hickey, “Reinforcement learning algorithms for robotic navigation in dynamic environments.,” ISA transactions, vol. 43, no. 2, pp. 217-30, Apr. 2004.

 

This is a good material for learning the reinforcement leanring algorithms in mobile riobotic control, This paper provides several ML angorithms, comparing them, and make simulations, and conclusions. Currently this is not my direction to go so that I won’t bother to spend so much time on it, but it is really good for reinforcement learning researchers and applications.

 

Note: Probabilistic taget detection by camera-equipted UAVs

Posted on by
Reply

Source:

[1] A. Symington, S. Waharte, S. Julier, and N. Trigoni, “Probabilistic target detection by camera-equipped uavs,” in Robotics and Automation (ICRA), 2010 IEEE International Conference on, 2010, vol. 67, pp. 4076–4081.

This is another example using camera information to track the target by UAV. In this paper the recursive bayesian estimator has been adopted in order to track the target in each frame data acquired  from camera, and that’s why it has something related with the machine learning.

Note: Learning to fly like a bird

Posted on by
Reply

Source:

[1] R. Tedrake, Z. Jackowski, R. Cory, J. W. Roberts, and W. Hoburg, “Learning to Fly like a Bird,” Under review, 2009.

This is an interesting paper. Learing to fly like a bird is really hard according to us engineers because the dynamics of a flying bird is so difficult and the control system of a bird is so elabrate. However I believe this paper gives a good starting point. This paper gives a detailed research on the principle of bird flying, its dynamics, the fluid dynamics, how to modelling, and some other control related problems. Forgive me my poor foundation on mathematics and control theory, I can hardly understand the details provided in this paper, but as desclared by this paper, that it uses a machine learning algorithms to get the approximate model of the aereal vehicle, a model based feedback design and modelless policy selection. That’s how machine learning be used in this application, but to know how I need to go deep into this paper to find out.

Note: Controller design for quadrotor UAVs using reinforcement learning

Posted on by
Reply

[1] H. Bou-Ammar, H. Voos, and W. Ertel, “Controller design for quadrotor UAVs using reinforcement learning,” in Control Applications (CCA), 2010 IEEE International Conference on, 2010, pp. 2130–2135.

This paper first  modeled for the quadrotor UAV and designed a non-linear controller, by uncoupling the control into two control loops: the velocity control and the altitude control. The sencond part of this paper used a reinforment learning algorithms to let the controller generate the optimized behavior. It used a Markov decision process to train the controller, together with some other details. The simulation and experiment results were also available in this paper.

Talking about the reinforcement learning, I think it is a kind of supervised control. It needs the support from lower level / real-time controller. The lower-level controller generates the action, the reinforcement learning controller judges this action, and adjust the command it has send to the lower-level controller to generate a more promising action / actions.

Note: Combined optic-flow and stereo-based navigation of urban canyons for a UAV

Posted on by
Reply

Source:

[1] S. Hrabar, G. S. Sukhatme, P. Corke, K. Usher, and J. Roberts, “Combined optic-flow and stereo-based navigation of urban canyons for a UAV,” in 2005 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2005, pp. 3309-3316.

This is a very good paper which combines the optic-flow technique and the stereo vision to guid the UAV navigating through the canyons above the urban street. This paper gives a good introduction on optic-flow technique for determining the relative distance of objects using moving camaras. With the introduction of this paper, the optic-flow is used to make the UAV flying between the two walls of the lane, and the stereo vision is used to detect the obstacle ahead and make a decision to turn left or right. The experiments shown in this paper is very encouraging and promising. I think it will be very useful in robotic rescue and search applications.

Note: Autonomous altitude estimation of a UAV using a single onboard camera

Posted on by
Reply

Source:

[1] A. Cherian, J. Andersh, V. Morellas, N. Papanikolopoulos, and B. Mettler, “Autonomous altitude estimation of a UAV using a single onboard camera,” in 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2009, pp. 3900-3905.

It is easily can be seen that this paper deals with the problem on how to deduce the altitude of the UAV from the image sequences acquired from the camera mounted on it. The author claimed it was the first time the problem had been addressed and settled. The basic idea is using the texture features, which may be not accurate because I didn’t go deep into this paper. The method is using the Markov Random Field (and Maximum A Posteriori), together with the supervised machine learning algorithms. It is may be useful for the autonomous control of an UAV during search and rescue activities.

Note: A vision based forced landing site selection system for an autonomous UAV

Posted on by
Reply

I have just finished looking through the paper titled ” a vision based forced landing site selection system for an autonomous UAV”, and here is the comments  I had made for this paper.

First, this paper defined the necessity of the forced/emergent landing site selection for an autonomous UAV. Although through GPS and existing map or GIS database it is would be possible for UAV to find a landing space, however it would be greatly restricted if there is no further information about some moving objects on the surface of landing site, or failed functioning when in disaster situation or some places without GIS data. The best solution is to mimic the man driving procedure, and find the landing site using camera mounted on the UAV.

Second, this paper provied a framework for finding the landing site using the video data, which has the layered structure.  It also put forward the functions of the first 4 layers of this framework, allowing adding other function layers to this framework without changing too much the whole system. Its working principle is from coarse to fine, that is: using line/edge recognition algorithms to quickly find the places which are possible suitable for landing, and then on layer 2 to determine the surface type of each cadidate landing site; the third layer is the slope augmentation so that the unsuitable landing sites will be rejected, and the 4th layer is the dangerous landing site identification. Finally, through all the layers, there will be a fusion mechanism to determine the best landing site for the UAV.

Third, the paper gave the configuation of the experiment rig and also, the testing results. The paper has done a good job according to the experiement data and figures, but no further information has been provided by this paper, and may need to do the same experiment again to see if it is really useful.

Last, the paper prived some suggestions on how to identify the surface type of the candidate landing sites. It listed many references, and would be helpful if i want to do the same job.

美国访学日记（3）

Posted on by
Reply

终于到了美国。

对于从来没有踏出过国门的我来说，还是有点激动的。

一下飞机，首先就是办理入境手续。由于一整飞机的人都要办理入境手续（也可能还有其他同时差不多到达的飞机吧，反正人挺多的)，办理入境手续的人非常多，需要排队，挺长的队伍。还好因为已经落地了，心里也不着急了。我们下一班的飞机要10个小时以后，所以就慢慢等着吧。入境手续分为美国公民和非美国公民，是不同的队伍。令人惊讶的是，我们在飞机上看到的好几个国内的老头老太太，居然走的是美国公民通道，也许拿到了绿卡了？

Continue reading →