af Graeme Smith 16 år siden
1607
Mere som dette
Draw the conclusions from my work. What have I shown? What can I say about using the uDS as the basis for radar ATR? What are the differneces between mono and multiperspective classification? What is the best way to select training data for the classifier?
Start to look at where the work can progress from here. What are the next steps that must be taken? How would it be best to take them?
Review my work in a critical manner. What are its strengths and what are its limitations? It is important here to keep the strengths up, it will be easy to tear my work apart because I am so close to it.
A comparison of the difference between monoperspective and multiperspective classification systems.
I expect the multiperspective classification to come out as the best. However this will need to be fully justified here.
Review the multiperspective classification work. What has it shown?
How does it contrast with the exisiting body of monoperspective uDS classification, my own monoperspective work and other methods?
Results will be merged into this section so that it is not necessary to flick backward and forward between the method and results sections.
Review the monoperspective frequency domain work.
How has it compared with the existing work in the same field?
What's been good about it and what's been not so good?
How do the theory/simulated results match with the actual data?
Review the monoperspective time domain work.
How has it compared with the existing work in the same field?
What's been good about it and what's been not so good?
How do the theory/simulated results match with the actual data?
Review the ideas of moving to a multiperspective classification system. What are the percieved gains? How do the techniques employed need to be modified?
Results will be merged into this section so that it is not necessary to flick backward and forward between the method and results sections.
How the work from the monoperspective case was modified to work with multiple perspectives.
How does feature extraction need to be modified in the multiperspective domain?
Describe how the models of the uDS were expanded into the multistatic domain.
Comparable data analysis to the monoperspective case. Looking at how things are different in the multiperspective domain. This will no doubt include elements of data fusion.
There will need to be some stuff in here about how to create some form of mulitperspective data from the Thales data i.e. temporally separated data.
What was considered ahead of doing the trials? How did the theory compare with the experimental results?
Keep this focused on issues of multistatic uDS and ATR. Don't spend ages talking about reflectivity coefficients and soil moisture.
Develop the idea of the multistatic μDS. The size of this section could depend on what happens in the μDS section of the Background chapter.
The basic idea is to consider what is actually meant by multistatic micro-Doppler. The focus will be on how to fuse the differnent bistatic channels together.
Review the different forms that the μDS can take. Effectively this is sensor fusion between the different bistatic paths that are available.
This should be tied quite closely to the NetRAD output. I'm not doing a multistatic radar thesis so I only want to focus on what I'm using and not trying to develop some all encompasing theory.
Multidimensional
Concatenate
The differences between monostatic, bistatic and multistatic radar. The most famous version of all the equations are monostatic. How do they vary in the bi/multistatic case?
Describe the template matching techniques that I have employed: DTW and k-NN. This introduction will cover the tequniques themselves and how I have employed them in classifying radar data.
Need to give some prominance to the use of DTW since this is the first time it's been used in radar.
Will also cover the creation of the libraries and test dataset here. This work may need to be moved to a position of more significance in the section.
The specifics for the datasets used in this round of testing. How did they differ from the ideal cases?
Principal component analysis
Freqeuncy normalization
Removing clutter
Zero Doppler clutterremoval
CLEAN algorithm
Describe how to create synthetic unknown data. Introduce the propeller model and discuss it's outputs.
This is a very short introduction to the chapter explaining what it's all about. I.e. we've got all this lovely data from Thales, how do we use it in classification testing?
Give an analysis of the data. What information can be found within it? What are the limits to this data?
Discuss things such as how the the uDS will vary. For example some graphs showing how the main Doppler shift and signature bandwidth will vary with target velocity.
Disucess the techniques used to observe the uDS. Look at what's good and what's bad about each one. Where would you want to use it? What assumptions are being made when it is used.
The focus will be on FFT and STFT/Spectrogram should probably find at least one other to use also.
It might be worth making the new approach cepstral processing as this is common in some other research on uDS and speech processing.
Consider cepstral processing for this and some of the more advanced techniques such as Wigner-Ville distribution. Why haven’t I used them?
Discuss the uDS theory. Run through the ideas of Chen and Schnieder on how the uDS may be described theoretically.
Attempt to do this conceptually using ideas such as the superposition of the returns from the sub components of the targets. This will help with the ideas of how to model a walking person that come later.
Initially I discuss the theory from a monostatic perspective. The multistatic situations will be included in it's own section. I must remember to leave signposts for the multistatic situation as I work through the monostatic theory.
How to evaluate the classifiers in the rest of the study.
This work is basically completed already, but there is the tricky issue of the Can't Declare thresholding. This is important, but I haven't implemented it in any work yet. Doing this is important as it will distinguish my analysis from Michele Vespe's.
Introduce the parameters used for evaluating a classifier.
Explain how they are related and how they are effected by the two thresholds (unknown and can't declare).
Describe the use of ROC curves to set the thresholds for unknown intput detection and can't declare.
Introduce the idea of can't declare and explain how it is different to unknown.
Review the importance of being able to dected unknown inputs.
Present the idea of the black box classifier. It is a high level abstraction that can be used to develop concepts without worrying about the precise implementation.
The literature survey. Perform a critical review of the current state of ATR and uDS classification.
What has been done on multiperspective classification?
The case for why my research is needed. What is it about the other methods, both the alternatives to the μDS and the existing μDS techniques, that are limitted?
Things to consider here are: the platform depent nature to the Thales system; mistakes made by tired human operators; the need for specialist radar for SAR/ISAR/HRR.
Review the current methods for using μDS to classify targets. Look at systems such as the Thales MSTAR and trained operator approaches.
Review the alternative methods to μDS target recognition.
Consider things such as SAR/ISAR imaging, HRR classification and JEM modulation techniques. What is the deal with RCS based classification?
Need to run through, at a high level, their underlying workings and what sort of successes they have.
A quick review of the basics of radar: the usual stuff that gets covered in all radar text.
Focus should be given to the concepts that will be used in the rest of the the thesis e.g. introduce things like SAR/ISAR that provide alternatives to uDS classification. Keep these discussions very brief.
Introduce the idea of the Doppler shift. Run into a little more detail than one would normally do on this. Doppler shift is essential to the work in hand so best explain it well. Get in ideas such as how it can be considered a changing phase etc.
In this section consider both the mono and multiperspecitve situation. Ideas is going to be that the multiperspective stuff is brought in a little later, but I need to set it up here and put in the signposts
Introduce the idea of pulse compression here. Discuss the way correlation can be used to generate a compressed pulse and lead into the ambiguity diagram. Place emphasis on the parts of the ambiguity function that relate to Doppler.
What is the Doppler shift?
How was it orginally detected? This is MTI.
How can it be detected now we have pulse Doppler radars?
What does it mean to have a coherent radar? How are the achieved?
Discuss what it is that make having a coherent radar important: integration and pulse Doppler.
Discuss the concepts of radar range equation, SNR, pulse ambiguity, and pattern propagation factor.
Introduce the basic compoents of a radar system. Use a block diagram and make sure that it is clear which part is the receive chain and which is the transmission ready for split into bi and multistatic.
Include a target in the diagram so that all aspects of the discussion can be related back to the diagram.
Introduce the subject of the PhD.
Acts a little like and extended abstract/executive summary.
Covers similar material to the background but at a much higher level.
I want to use this section to orientate the reader, start bringing in the terminology of the subject and the basic ideas.
It will be high-level and should be a complete description of the work. A busy person should be able to read the introduction and the conclusion as a stand alone document.
Describe the structure of the thesis to the reader.
A note on the emerging field. Probably should have some kind of reference to Michele's work and perhaps to the bat paper that covers the temporally separate multiperspective approach.
Completely novel. The nearest thing to using multistatic radar and uDS is the paper on bistatic helicopter signature. But bistatic and multistatic are very different things.
Depends on an outstanding activity to compare the correlation libraries with randomly selected ones.
The contrasting will be helped when I get away from k-NN that needs all the extra entries in it's library.
However, whatever the output I should be able to say something in either why random selection or controlled selection is best. This hasn't really been done before.
A note on applying dynamic time warping to radar data. The other techniques I shall apply have all been used in the radar context before. However, to the best of my knowledge DTW hasn't been used with radar data previously.
How has my worked moved the field of uDS recognition for radar ATR forward? How is why I have done different from the work of Stove, Bilik & Janhangir?
Give a brief introduction to the ideas of radar. Rapidly move onto the idea that with recent advances in computing technolgoy it has become practical to add target reconition to radar's traditional roles of detection and ranging.
