Please use this identifier to cite or link to this item:
|Title:||Sparse representation and its applications to airborne platforms||Authors:||Srinivasan, Abirami||Keywords:||Engineering::Electrical and electronic engineering::Electronic systems::Signal processing||Issue Date:||2021||Publisher:||Nanyang Technological University||Source:||Srinivasan, A. (2021). Sparse representation and its applications to airborne platforms. Doctoral thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/148844||Abstract:||Many natural signals can be concisely represented in a different domain. These signals, are then said to be sparsely represented. Approximation of a signal with an overcomplete system gives a sparse representation of the signal based upon redundant “dictionaries”. The replacement of the original signal with a sparse representation does not only result in data compression, but also provides an alternative look into the inherent structure of the natural signals with the overcomplete dictionaries. Sparse representation finds application in various fields such as signal processing, image processing, pattern recognition, etc. Numerous algorithms are increasingly being proposed for sparse representation. The advantages of sparse representation can be exploited well for processing associated with signals generated on airborne platforms. The work discussed in this thesis explores the use of sparse representation in signal processing related to airborne platforms such as drones, helicopters and Synthetic Aperture Radar (SAR). Compressed Sensing is a direct relative of sparse representation, and has grown increasingly popular in the recent years. Compressed Sensing (CS) has been emerging as an effective sampling scheme for signal acquisition and efficient sensor design which not only achieves data compression but also enables various forms of signal detection and estimation with vastly reduced data. CS relies on the inherent “sparsity”, or compressibility of a signal in some basis and operates on far fewer samples than traditional sampling schemes. The signals, thus sensed, can be reconstructed accurately, with high probability, using convex optimization techniques such as l1-optimization or more efficient greedy recovery algorithms such as Orthogonal Matching Pursuit. This new sampling scheme finds applications in various areas of signal processing. The application of sparse representation-based dictionary in anomaly detection of accelerometer data obtained from helicopters is explored in this thesis. Sparse-representation based anomaly detection uses the sparsity of the signals to train a dictionary. A Sequential Generalization of K-means (SGK), a dictionary training algorithm is used towards this end. The training data which are classified “normal”, typically contains a wide set of signals of which some show significant deviation from the more common statistical features of the rest. In such a scenario, a dictionary is required to be retrained to be able to represent the signals reasonably across the sample space. A modification to the SGK, the “Weighted Sequential Generalization of K-Means” is proposed here to be able to represent the signals more sparsely. The performance of the Weighted SGK and the SGK in retraining are compared with the validation set of the data. It is observed that Weighted SGK outperforms SGK in the detection performance for the anomalies in the helicopter data. Weighted SGK is also seen to outperform some classical machine learning approaches. Synthetic Aperture Radar (SAR), that is used in remote sensing for high resolution imagery by utilizing flight path to synthesize a large aperture, generates large amount of data that are transmitted to the ground station where it is further processed. CS exhibits the potential to reduce this data that needs to be transmitted through the downlink. CS-based Low Rank and Sparse Decomposition (LRSD) is proposed as a solution to tackle the data rate on the downlink. Proposed LRSD has two-fold advantage of compression of raw data as well as useful image information isolated in one of the two components, either the low-rank or the sparse. The quality of the images are compared among those obtained by the uncompressed data, CS-compressed data and CS-compressed-LRSD-decomposed data. It is shown that the metrics used to indicate certain visual aspects of the images gave reasonable results for the CS-based LRSD scheme, as compared to the uncompressed and the traditional CS-compressed data. The use of inherent nature of the phase in SAR signals is a rather unexplored area and is of potential interest. Surveillance systems are widely used not only for military but also for everyday civilian purposes. Intrusion-detection schemes are of interest in every surveillance system. Mobile flying platforms such as drones are being used effectively as an autonomous surveillance system that can not only detect an intrusion on-board, but also make an informed decision to either track the intruder or transmit the required information to the ground station. Efficient techniques to implement a surveillance system on-board continue to be explored. A CS-based scheme for intrusion detection on-board the AR Parrot Drone is implemented. It is observed that a CS-based system can make efficient use of on-board resources by exploiting the properties of the compressed data space. It is also capable of efficient usage of the available bandwidth by transmitting frames to the ground station only if an intrusion is detected. Performance metrics, namely, encoding time and detection accuracy were evaluated. The detection accuracy across data sets that were tested was found to be excellent. The encoding time using CS on-board outperformed JPEG encoding time for similar region-of-interests (ROIs).||URI:||https://hdl.handle.net/10356/148844||DOI:||10.32657/10356/148844||Rights:||This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Updated on Jan 21, 2022
Updated on Jan 21, 2022
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.