Anomaly detection & object classification using multi-spectral LiDAR and sonar
Chhabra, Puneet Singh
MetadataShow full item record
In this thesis, we present the theory of high-dimensional signal approximation of multifrequency signals. We also present both linear and non-linear compressive sensing (CS) algorithms that generate encoded representations of time-correlated single photon counting (TCSPC) light detection and ranging (LiDAR) data, side-scan sonar (SSS) and synthetic aperture sonar (SAS). The main contributions of this thesis are summarised as follows: 1. Research is carried out studying full-waveform (FW) LiDARs, in particular, the TCSPC data, capture, storage and processing. 2. FW-LiDARs are capable of capturing large quantities of photon-counting data in real-time. However, the real-time processing of the raw LiDAR waveforms hasn’t been widely exploited. This thesis answers some of the fundamental questions: • can semantic information be extracted and encoded from raw multi-spectral FW-LiDAR signals? • can these encoded representations then be used for object segmentation and classification? 3. Research is carried out into signal approximation and compressive sensing techniques, its limitations and the application domains. 4. Research is also carried out in 3D point cloud processing, combining geometric features with material spectra (spectral-depth representation), for object segmentation and classification. 5. Extensive experiments have been carried out with publicly available datasets, e.g. the Washington RGB Image and Depth (RGB-D) dataset , YaleB face dataset1 , real-world multi-frequency aerial laser scans (ALS)2 and an underwater multifrequency (16 wavelengths) TCSPC dataset collected using custom-build targets especially for this thesis. 6. The multi-spectral measurements were made underwater on targets with different shapes and materials. A novel spectral-depth representation is presented with strong discrimination characteristics on target signatures. Several custom-made and realistically scaled exemplars with known and unknown targets have been investigated using a multi-spectral single photon counting LiDAR system. 7. In this work, we also present a new approach to peak modelling and classification for waveform enabled LiDAR systems. Not all existing approaches perform peak modelling and classification simultaneously in real-time. This was tested on both simulated waveform enabled LiDAR data and real ALS data2 . This PhD also led to an industrial secondment at Carbomap, Edinburgh, where some of the waveform modelling algorithms were implemented in C++ and CUDA for Nvidia TX1 boards for real-time performance. 1http://vision.ucsd.edu/~leekc/ExtYaleDatabase/ 2This dataset was captured in collaboration with Carbomap Ltd. Edinburgh, UK. The data was collected during one of the trials in Austria using commercial-off-the-shelf (COTS) sensors.