Ripul Ghosh

Moving target detection in thermal infrared imagery using spatiotemporal information

An efficient target detection algorithm for detecting moving targets in infrared imagery using spatiotemporal information is presented. The output of the spatial processing serves as input to the temporal stage in a layered manner. The spatial information is obtained using joint space-spatial-frequency distribution and Rényi entropy. Temporal information is incorporated using background subtraction. By utilizing both spatial and temporal information, it is observed that the proposed method can achieve both high detection and a low false-alarm rate. The method is validated with experimentally generated data consisting of a variety of moving targets. Experimental results demonstrate a high value of F-measure for the proposed algorithm.

Thermal Imaging And Its Application In Defence Systems

Thermal imaging is a boon to the armed forces namely army, navy and airforce because of its day night working capability and ability to perform well in all weather conditions. Thermal detectors capture the infrared radiation emitted by all objects above absolute zero temperature. The temperature variations of the captured scene are represented as a thermogram. With the advent of infrared detector technology, the bulky cooled thermal detectors having moving parts and demanding cryogenic temperatures have transformed into small and less expensive uncooled microbolometers having no moving parts, thereby making systems more rugged requiring less maintenance. Thermal imaging due to its various advantages has a large number of applications in military and defence. It is popularly used by the army and navy for border surveillance and law enforcement. It is also used in ship collision avoidance and guidance systems. In the aviation industry it has greatly mitigated the risks of flying in low light and night condi...

Optical fiber antenna generating spiral beam shapes

A simple method is proposed here to generate vortex beam and spiral intensity patterns from a Gaussian source. It uses a special type of optical fiber antenna of aperture ∼80 nm having naturally grown surface curvature along its length. The antenna converts linearly polarized Gaussian beam into a beam with spiral intensity patterns. The experimentally obtained spiral patterns with single and double spiral arms manifest the orbital angular momentum, l = ±1, 2, carried by the output beam. Such beam can be very useful for optical tweezer, metal machining, and similar applications.

Computational techniques for classification of military vehicles using seismic signatures

In this research work a seismic classification system is designed to distinguish between tracked and wheeled vehicle classes. Owing to the extreme non-stationary nature of seismic signals, choosing robust features is an important aspect for the purpose of classification. To obtain a varied feature set different signal processing techniques namely Fast Fourier Transform (FFT), Walsh-Hadamard Transform (WHT), Hilbert-Huang Transform (HHT) and Wavelet Transform (WT) are investigated. Dominant features are identified from the feature bank using Principal Component Analysis (PCA). This choice of optimal and robust features leads to a better class discrimination. It is observed that the classification results obtained by the varied feature set followed by optimization has improved classification accuracy of 95% than using features extracted from individual signal processing techniques.

Target Recognition in Infrared Imagery Using Convolutional Neural Network

In this paper, deep learning based approach is advocated for automatic recognition of civilian targets in thermal infrared images. High variability of target signature and low contrast ratio of targets to background makes the task of target recognition in infrared images challenging, demanding robust adaptable methods capable of capturing these variations. As opposed to the traditional shallow learning approaches which rely on hand engineered feature extraction, deep learning based approaches use environmental knowledge to learn and extract the features automatically. We present convolutional neural network (CNN) based deep learning framework for automatic recognition of civilian targets in infrared images. The performance evaluation is carried on infrared target clips obtained from ‘CSIR-CSIO moving object thermal infrared imagery dataset’. The task involves four categories classification one category representing the background and three categories of targets -ambassador, auto and pedestrians. The proposed CNN framework provides classification accuracy of 88.15 % with all four categories and 98.24 % with only three target categories.

Time‐Frequency Approach for Stochastic Signal Detection

The detection of events in a stochastic signal has been a subject of great interest. One of the oldest signal processing technique, Fourier Transform of a signal contains information regarding frequency content, but it cannot resolve the exact onset of changes in the frequency, all temporal information is contained in the phase of the transform. On the other hand, Spectrogram is better able to resolve temporal evolution of frequency content, but has a trade‐off in time resolution versus frequency resolution in accordance with the uncertainty principle. Therefore, time‐frequency representations are considered for energetic characterisation of the non‐stationary signals. Wigner Ville Distribution (WVD) is the most prominent quadratic time‐frequency signal representation and used for analysing frequency variations in signals.WVD allows for instantaneous frequency estimation at each data point, for a typical temporal resolution of fractions of a second. This paper through simulations describes the way time frequency models are applied for the detection of event in a stochastic signal.

A spatio-temporal deep learning approach for human action recognition in infrared videos

Human action recognition in indoor environment can prove to be very crucial in avoiding serious accidents and (or) damage. Application domain spans from monitoring the actions of solitary elders or persons with disabilities to monitoring persons working alone in a chamber or in isolated industry environment. These scenarios demand an automatic near real-time activity recognition and alert to save life and assets. In this work, considering the fact that the sensing modality should be capable of working round the clock in a non-intrusive manner, we have opted for thermal infrared camera, which captures the heat emitted by objects in the scene and generates an image. Motivated by the recent success of convolutional neural networks (CNN) for human action recognition in IR images, we extend this work by incorporating one additional dimension i.e. the temporal information. In this work, we have designed and implemented a 3D-CNN for learning the spatial as well as the sequential features in the thermal IR videos. In this work, eight action classes are considered - Walking, Standing, Falling, Lying, Sitting, Falling from chair, Sitting up (recovering from fall from sitting posture), Getting up (recovering from fall from lying posture). To evaluate the proposed framework, infrared (IR) videos of different actions were generated in three diverse environments of home – inside study room, inside a bedroom and in the garden. The dataset comprised of 2641 and 894 IR videos for training and testing respectively, each of half a second duration performed by more than 50 volunteers. We have designed and implemented 3D-CNN, comprising of two blocks, each of two convolution and one max pool layer, which automatically constructs features from raw data incorporating both spatial and temporal information to learn actions. Network parameters are learned using back-propagation algorithm and the learning is supervised. Experimental results indicate 85% classification accuracy on 894 complex test videos of the proposed Spatio-Temporal Deep Learning architecture on the IR action dataset.

Towards an optimal bag-of-features representation for vehicle type classification in thermal infrared imagery

Automatic vehicle type classification plays a significant role in security, traffic control and autonomous driving applications. Thermal infrared (IR) cameras operating even in complete darkness and adverse weather conditions emerge as a potential sensing modality for such challenging outdoor applications. However, automated vehicle type classification in infrared imagery still poses significant challenges due to high variability of vehicle signature in infrared band leading to high intra-class variation and low inter-class variation. To address these issues, we demonstrate the use of local features represented in a bag of words framework. In this work, we present comparative analysis of two feature detectors, MSER – a sparse region based detector and uniform dense sampling of points in the image across multiple scales (termed dense). A bag of features (BoF) framework based on SURF feature descriptor and SVM classifier for vehicle type classification are evaluated on a thermal infrared (TIR) vehicle dataset. A number of variations are present in the TIR vehicle dataset - scale variation, pose variation and partial visibility of vehicles captured under varied environmental conditions. The dataset contains four vehicle categories commonly plying on Indian roads, Bike, Autorickshaw, Car and Heavy vehicle. The performance of the designed vehicle type classification framework was evaluated using performance metrics, classification accuracy and confusion matrix. The optimized sparse MSER and dense BoF framework demonstrated decent classification accuracies of 85.7% and 93% respectively for automatic vehicle type classification on the thermal infrared vehicle dataset.

Machine Learning Based Comparative Analysis for the Classification of Earthquake Signals

This research aims at classifying earthquake signals from seismic noises caused due to anthropogenic activities. We aim at designing a seismic classifier for classifying true earthquake signals so as to reduce the false alarms thereby avoiding excessive data logging due to cultural noise. Based on theoretical and experimental consideration, a set of time and frequency domain features are extracted and used as features to train the supervised classifier network, viz., k-nearest neighbor (k-NN), maximum likelihood (ML), artificial neural network (ANN), and support vector machine (SVM). Two datasets were used in this research work K-NET (Kyoshin Network), Japan and strong motion seismic data recorded at CSIR-CSIO, Chandigarh using BASALT accelerograph of Kinemetrics Inc. Comparative analysis of the classifiers shows that SVM outperforms the other methods with an accuracy of 99.60%.

Deep learning approach for human action recognition in infrared images

Abstract Human action recognition based Ambient assisted living (AAL) systems, targeted towards providing assistance for the elderly and persons with disabilities, have been of interest to researchers from various disciplines. The research primarily focuses on development of automatic, minimally intrusive and privacy preserving systems. Although popular in the strategic sector, thermal infrared (IR) cameras haven’t been explored much in AAL. This work demonstrates the use of IR cameras in the field of AAL and discusses its performance in human action recognition (HAR). Particular attention is drawn towards one of the most critical actions - falling. In this reference, a dataset of IR images was generated comprising of 6 action classes – walking, standing, sitting on a chair, sitting on a chair with a desk in front, fallen on the desk in front and fallen/lying on the ground. The dataset comprises of 5278 image samples which have been randomly sampled from thermal videos, each of about 30 s, representing the six action classes. To achieve robust action recognition, we have designed the supervised Convolution Neural Network (CNN) architecture with two convolution layers to classify the 6 action classes. Classification accuracy of 87.44% has been achieved on the manually selected complex test data.