Mukesh A. Zaveri

Research Interest

Computer Vision, Image Processing
Audio-Speech Signal Processing
Machine Learning, Wireless Sensor Network
Internet of Things

Contact Information

Computer Engineering Department
SVNIT, Surat, Gujarat, India
Pin code: 395007
Office phone: (0261) 220 1766
Mobile: +91-9427581925

Professional Experience

Ph.D. Research Work

Ph.D. Thesis Title: Detection and Tracking of Point Targets in IR Video Sequence

Under the guidance of Prof. Uday B. Desai and Prof. S. N. Merchant, Electrical Engineering Department, IIT Bombay

(This research project was funded by Department of Extramural Research and Intellectual Property Rights, (IRDE Lab) DRDO, Govt. of India, New Delhi.)

In this thesis the algorithms are introduced to detect and track targets in an air-borne infrared image sequences. The algorithms are mainly developed for detection and tracking of multiple point targets without using any apriori information about the target dynamics. Generally, the movement of the targets is arbitrary. The tracking algorithms must be able to track maneuvering and non-maneuvering targets simultaneously. The detection of point target is very challenging task due to lack of any texture information. In such a case, the motion is used as a cue and the detection algorithm is proposed which uses the wavelet transform for temporal multiscale decomposition. The algorithm is further extended for detecting the approaching targets in infrared image sequences.

The detection is followed by the tracking of multiple point targets in the presence of dense clutter. Two important factors for success of any tracking algorithms are; the selection of the dynamic model used for tracking and the data association method used for observation to track assignment. The data association is required for updating the target hidden state parameters. In the thesis various tracking algorithms are proposed targeting mainly the following issues: (i) tracking arbitrary movement of the targets with minimal number of models (ii) robust and efficient data association method and, (iii) the filtering method used for tracking.

Data association is crucial in the presence of multiple targets and clutter. In the current research work, the various methods are explored (a) based on an implicit observation to track assignment, and (b) the methods which evaluates the assignment weights for each observation and subsequently, these weights are used for state update. The use of assignment weights of data association avoids the uncertainty about the origin of an observation. The later method uses all observations, validated by means of statistical minimum distance criterion, for state update and does not assign an observation to any track implicitly. In this context, various data association methods based on (a) implicit observation to track assignment, i.e. nearest neighbor method, using state vector of the model and the genetic algorithm, (b) assignment weights for each observation using Expectation Maximization algorithm, neural network, Markov random field, and the genetic algorithm, have been proposed.

The proposed target detection and tracking algorithms are as follows:

Worked on Research Projects

Membership to Technical Society

Little "Star" Achievement - Prizes & Awards