Academic Background- October'2018-Present: Master of Informatics in Department of Intelligence Science and Technology, Kyoto University, JAPAN</ul>
- Research Laboratory: Speech and Audio Processing, Kyoto University, JAPAN</ul>
- May'2017-August'2018: Project Research Assistant, Electrical Engineering, IIT Bombay, INDIA</ul>
- 2013-17: B.Tech, Aerospace Engineering, IIT Bombay, INDIA</ul>
Technical Skills- Programming Language: Python, Matlab, C/C++, HTML
- Libraries: Tensorflow, Keras, Scikit-learn, OpenCV
- Software Tools: MS Office, ANSYS Fluent, AutoCAD
Research Interest- Machine Learning, Signal Processing, Sequential data modeling, Speech Recognition using deep learning techniques, Image classification and object recognition, Natural Language Processing</ul></li>Publication
- 'Acoustic Localization of Gunshots in the Presence of Obstacles'</strong> with Aniruddha Sinha, General Acoustics, AIAA Aeroacoustics Conference, Atlanta, Georgia, USA, 26 June 2018 (View pdf)</ul>
- 'Automatic Classification of Acoustic Signals from Gunshots' with Sourav Bhattacharya and Aniruddha Sinha, Proceedings of the 7th International and 45th National Conference on Fluid Mechanics and Fluid Power (FMFP)
December 10-12, 2018, IIT Bombay, Mumbai, India (View pdf)</ul>
Curriculum Vitae- View pdf</ul>
</td>
About me
I'm Tushar, Graduate Research Student of Graduate school of infomatics in Department of Intelligence Science and Technology specialized in Automatic Speech Recognition advised by Prof. Tatsuya Kawahara at Kyoto University, JAPAN. My primary field of research interest is Artificial Intelligence (AI).
Within AI, I am interested in problems related to probabilistic modeling, machine learning, and their interdisciplinary applications to engineering and biomedical fields. Motivated by the emergence of large datasets in various areas of science & technology, business, and everyday life that drives me to focused on machine learning and deep learning that aims to improve the understanding of patterns in the data and processes that generate them, and utilize them in solving nontrivial and challenging problems
</span>
| </tr>
</tbody>
</table>
</br >
Research Project | Gunshot Localization Technology
Sponsored by National Center of Excellence in Technologies for Internal Security, IIT Bombay
Gunshot Event Detetcion (Glock Pistol)
Muzzle blast gunshot event is detected using constant size sliding window technique. Recording was taken at frequncy 50Hz with "Sennheiser MD42" dynamic microphone. Person with Glock pistol standing at 6.4 meters from microphone and angle between line of firing & microphone is 48 degree in clockwise direction. Standard deviation of signals values was computed and thresholded at previous window and placed the condition on current and previous window. The condition is propagated through window with time to appropriately get the time indices of events.
Gunshot Event Detection (AK-47 Rifle)
AK-47 bullet travels more than the speed of sound since shockwave can be captured along with muzzle blast. Both events is detected using constant size sliding window technique. Recording was taken at frequency 50Hz with "Sennheiser MD42" dynamic microphone. Person with AK-47 standing at 24.4 meters from microphone and angle between line of firing & microphone is 12 degree in clockwise direction. Standard deviation of signals values was computed and thresholded at previous window and placed the condition on current and previous window. The condition is propagated through window with time to appropriately get the time indices of events.
Extracted the acoustic signature of gunshot
4 Millisecond window of muzzle blast signature of glock pistol using event time indices and construct the dataset for pattern recognition and classification.
Elimination of ground reflections from gunshot signatures
The motivation to eliminate the ground reflections is to precisely localize the gunshot. The core idea for achieving this goal is to use autocorrelation to determine the time delay and amplitude ratio between direct and reflected signals. The reflected information is then filtered out using a finite impulse response filter. The model focuses on adjusting the parameters that define the amount of inhibition to the signal parameters in order to best reduce the reflections information.
Time Series acoustic classification with Tensorflow
Tensorflow framework is used for the implementation and training of the model for frame level acoustic classification using labeled extracted dataset of acoustic signatures of gunshots which contains the raw time-series data. 95% accuracy is achieved for the validation set.
Classification using CNN based on Spectrogram
Designed and trained convolutional neural network architecture using spectrogram as input to perform the sound classification task. Algorithm was performed on small dataset and achieved 100% accuracy in 21 epochs on test data.
Gunshot Localization
Acoustic waves resembling typical muzzle blasts are obtained in the far field by choosing a Gaussian pulse as the initial condition for pressure. Three clusters each contains 3 sensors is used to measure the pressure values. Contours of pressure at various time instants in free space
Gunshot Localization in the presence of small obstacle
Contours of pressure at various time instants with small square obstacle.
Gunshot Localization in the presence of large obstacle
Contours of pressure at various time instants with lagre rectangular obstacle.
Timeseries of pressure recorded at the second cluster (probe nos. 4, 5, and 6)
Timeseries of pressure recorded at the second cluster (probe nos. 4, 5, and 6) in the cases of (a)
free space, (b) small obstacle, and (c) large obstacle. The x−y coordinates of the probes are indicated.
Bearing angle errors (Scatter plots)
Bearing angle errors in case of source located at (0, 0) and sensor array centered at (48λ, 0) with
square obstacle of various sizes as indicated in the plot titles. In each plot, the error is represented both by
the marker size and the colour scale at (x/λ, y/λ) coordinates where the obstacles are separately centred.
Bearing angle errors (Line plots) Bearing angle estimation error (in degrees) for a point source located at (0, 0) and sensor arrays
centred at (48λ, 0), λ being the pulse wavelength. Each row corresponds to one of the four arrays described
in Table. Square obstacles of four sizes are tested as indicated at the top of each column. In each subplot,
the error is plotted vs. the obstacle centroid position (x) along the line joining the source to the sensor array.
The four curves in each subplot are for different normal-to-the-line distances of the obstacle center (y)
RELEVANT PROJECTS
Cluster Analysis
Identifying a unique people group on the basis of variation in name, date of birth and gender and hence deduplication of records was analysed using k-means clustering unsupervised machine learning algorithm to identify unique patients. Principal features is extracted using principal component analysis to visualize the groups of same people.
Dog breed image classification
Classification task was performed based on ensemble learning. Deep Learning models is implemented to classify the various classes of dog breed
COURSE WORK
|
Relevant theory courses
Data Analysis and Interpretation, Linear Algebra, Differential Equations (Basic), Computer Programming and Utilization, Calculus, Differential Equations (Advance), Signals and Feedback Systems, Introduction to Numerical Analysis, Introduction to Electrical and Electronics Circuits, Engineering Design Optimization.
Online Platform Courses:
Machine Learning A-Z: Hands-on Python (Udemy Inc. ), Deep learning A-Z: Hands-on Artificial Neural Network (Udemy Inc. ), Convolutional Neural Network (Stanford University)
|
</td> </tr>
</tbody>
</table>
|
