Moving Target Classification with Radar Point-Clouds and Supervised Contrastive Learning
(2021) In Master's Theses in Mathematical Scieces MASM02 20211Mathematical Statistics
- Abstract
- This thesis deals with radar data for the purpose of moving target classification in the context of surveillance. The radar data in question comes in the form of point-clouds represented as frame-wise histograms with several channels and we seek to improve upon an existing cross-entropy based deep learning classifier using supervised contrastive loss.
We find that the embeddings output by supervised contrastive loss exhibit clearer separation between the classes and that it is possible to combine this approach with principal component analysis to obtain greater classification scores. - Popular Abstract
- Technological advancements have enabled massive amounts of data to be collected. So much so that it has become virtually impossible to manually process more than a small fraction of all the information that is available.
In order to confront this challenge, we instead turn to algorithms to do the job for us. The choice of such algorithms is necessarily contingent upon context and purpose, but a powerful set of techniques fall under the umbrella term which is machine learning.
With machine learning, the goal is not to specify a deterministic algorithm for the end processing task directly but instead to implement a set of algorithms that can learn from ”experience”, where experience refers to available data.
In this thesis, we work... (More) - Technological advancements have enabled massive amounts of data to be collected. So much so that it has become virtually impossible to manually process more than a small fraction of all the information that is available.
In order to confront this challenge, we instead turn to algorithms to do the job for us. The choice of such algorithms is necessarily contingent upon context and purpose, but a powerful set of techniques fall under the umbrella term which is machine learning.
With machine learning, the goal is not to specify a deterministic algorithm for the end processing task directly but instead to implement a set of algorithms that can learn from ”experience”, where experience refers to available data.
In this thesis, we work with radar point-cloud data for the purpose of classification. The context is surveillance whereby a radar system is in mind which monitors a given area that needs to be protected from intrusion and theft. The idea is to use labeled data to train a machine learning algorithm to be able to discriminate between targets so that an alarm can be triggered when unauthorized subjects enter the surveilled area.
In order to accomplish this end, we take primary interest in the special category of machine learning models known as deep learning. These models require a great deal of data in order to be effective are and generally too complex to be fitted directly to the data. Hence they need to be fitted iteratively and approximately and the principal method used for this is stochastic gradient descent with cross-entropy loss.
We investigate the possibility of replacing the cross-entropy loss with so-called contrastive loss which is a recent idea in the field. With contrastive loss the employed deep learning model does not learn classification directly but rather the features corresponding to the involved classes. The model which learns the features is called the encoder network. Once the features have been learnt, any machine learning algorithm for classification can be coupled with the obtained encoder in order to do prediction.
The obtained results are promising in comparison to a provided baseline model in terms of various performance metrics such as accuracy, precision and recall. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9058570
- author
- Zandler Andersson, Nils LU
- supervisor
- organization
- alternative title
- Klassificering av rörliga mål med radardata och djupinlärning
- course
- MASM02 20211
- year
- 2021
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Statistical Learning, Machine Learning, Deep Learning, Radar, Radar Point-Clouds, Supervised Contrastive Learning
- publication/series
- Master's Theses in Mathematical Scieces
- report number
- LUNFMS-3098-2021
- ISSN
- 1404-6342
- other publication id
- 2021:E42
- language
- English
- id
- 9058570
- date added to LUP
- 2021-07-05 15:52:24
- date last changed
- 2021-09-01 03:41:09
@misc{9058570,
abstract = {{This thesis deals with radar data for the purpose of moving target classification in the context of surveillance. The radar data in question comes in the form of point-clouds represented as frame-wise histograms with several channels and we seek to improve upon an existing cross-entropy based deep learning classifier using supervised contrastive loss.
We find that the embeddings output by supervised contrastive loss exhibit clearer separation between the classes and that it is possible to combine this approach with principal component analysis to obtain greater classification scores.}},
author = {{Zandler Andersson, Nils}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Scieces}},
title = {{Moving Target Classification with Radar Point-Clouds and Supervised Contrastive Learning}},
year = {{2021}},
}