Over the course my graduate study, I have worked on several interesting projects for my courses. Here is a list of them.
Dynamic Social Grouping and Routing in Delay Tolerant Networks: (Mobile Sensor Data Management, Spring 2009)
Delay Tolerant Networks (DTNs) are characterized by intermittent connectivity, long and variable delays and asymmetric data rates. Mobility of the network nodes only adds difficulty to routing or data aggregation in these networks. As with any general purpose sensor network scenario, it is not always feasible to transmit directly to a sink node or base station. Also portions of the network may be disconnected for a long time and relaying or storing messages for the sink node may not be feasible due to limited capabilities of the nodes.
In this project, a new routing and data aggregation scheme exploiting, the underlying social network dynamics of the mobile sensor nodes. The proposed schemes also incorporates a dynamic grouping algorithm, using which nodes form groups dynamically over their life time, therefore, requiring no knowledge of the underlying network topology.
The proposed scheme is tested on a real world sensor network dataset and compared against epidemic routing, a peer-to-peer routing algorithm. Download the project report for Dynamic Social Grouping and Routing in Delay Tolerant Networks.
Parallel Image Registration using Graphical Processing Units: (Parallel Programming with MPI, Spring 2009)
Image registration, also called as Image Mosaicing or Image Stitching, involves the estimation of transformation from one image coordinate system to another. The transformation could include translation, rotation, shearing, scaling, and/or warping. The process consists of first estimating the translation between two images using phase correlation. Affine transformation matrix between the two images is then estimated using Sum of Absolute Differences (SAD) block feature tracking algorithm. Reed-Xiaoli (RX) algorithm was used to provide high quality features for tracking across two images.
The serial version of this tool was developed using C++/IPP (Intel Integrated Performance Primitives). CUDA (Compute Unified Device Architecture) was used to develop the GPU version of the tool. Since GPUs cannot perform branching efficiently, calculation and refinement of the Affine transform was performed on the CPU.
The serial version was run on a 3.2 GHz, Pentium Core2Duo system, running Fedora 10. The code used IPP v5.8 and was compiled using GCC 4.3. CUDA v1.0 was used to compile the GPU code and was run on a 3.2 GHz, Pentium Core2Duo system, running Windows XP. The code was compiled using Visual C++ 2005 compiler. Download the presentation for Parallel Image Registration Using Graphical Processing Units(.ppt.zip)
Course Project for Advanced Evolutionary Computing, Spring 2009.
Unpublished thesis material, description and report will be posted at a later date.
Evolving Web Personalization Mining: (Web Data Management and XML, Fall 2007)
Web Usage Mining is a technique of mining browsing behaviors from secondary data such as web server log, web proxy logs, content level logs and through the use of cookies or applets. Although this has been in wide use, the problem of user session discovery and clustering is still open for improvement.
A two phase PSO-k-means cluster optimization algorithm is developed to cluster user sessions from web server logs. The server logs are preprocessed to extract user sessions and relevant session information. Each session is represented as a set of vector representing weight of a particular page in the session. In phase one, Particle Swarm Optimization (PSO) algorithm is run for a predefined number of iterations. The output of this is a set of cluster, of which, clusters within a threshold distance of each other are merged. The resulting cluster centroid are supplied to k-means as input. The output of k-means algorithm is the user behavior clusters of interest. Download the project report for Evolving Web Personalization Mining.
Image Segmentation for Goal and Path Detection for Mobile Robots: (Digital Image Processing, Spring 2007)
This is part of a larger project for designing a leader-follower based multi-robot system. In this project, vision based goal detection and path detection is developed. The goal is defined as a red circle and is detection by applying OpenCV’s Freeman Method based contour detection algorithm. Path detection is achieved via image segmentation using a watershed algorithm. A morphological gradient of the acquired image is first calculated to increase the contrast of the edges. Watershed segmentation is then performed on the gradient image to achieve path detection. Download the project report for Image Segmentation for Goal and Path Detection for Mobile Robots.
Visualizing Rigid Body Collisions: (Java GUI and Visualization, Fall 2006)
Developed a Java based simulation for visualizing the dynamics of multiple circular bodies in a container, i.e., rigid body collisions without friction and gravity. The application depicts a rich user interface with dynamic menus and otehr interface elements. Download Java JAR file for Rigid Body Collisions. (Please note that this is written in Java 1.4 and therefore some functions may have deprecated. Opening this jar may cause your browser to become unresponsive or crash. Use caution.)
Face Recognition using Eigenfaces and Neural Network: (Introduction to Neural Networks and Applications, Fall 2006)
Designed a multi-layer feedforward backpropagation neural network for face recognition in image data. Principal Components Analysis (”Eigenfaces”) is applied on the image dataset is used for feature extraction. The resulting reduced dimensionality features is used to train the neural network. Image recognition is also performed in the PCA space. Effect of training algorithm, dimensionality reduction and image subsampling on the performance of the neural network was studied. Download the project report for Face Recognition using Eigenfaces and Neural Networks.
Dynamic Modeling of a Quadrupedal Robot with Bounding Gait: (Advanced Dynamics of Machines, Fall 2006)
Developed and simulated the dynamic model of a quadrupedal robot with bounding gait. The case of passive dynamic running was considered. Equations of motion were developed using Lagrange’s equations. Simulation was developed using Maple. Download the project report for Dynamic Modeling of a Quadrupedal Robot with Bounding Gait.
Hexapod Robot: (Robotic Manipulator and Mechanisms, Spring 2006)
Designed and Fabricated a six legged robot, and its associated gait mechanism and control system.
Other Projects
Learning Applied to Ground Robotics (LAGR):
Worked as a member of the UMR LAGR team. Served in the capacities of a programmer and system administrator. The project was aimed at developing a learning architecture for vision-based navigation for autonomous ground robots. The team developed a novel representation of visual information using a feature-flow model incorporating a view-centric approach. We also developed and implemented algorithms for vision based perception and navigational planning for off-road navigation.
Using Open Source Software for Real-Time Geospatial Applications – A Feasibility Study:
Conducted a survey on current open source geospatial solutions to assess their ability to be part of a real-time web-based geospatial analysis system. Tested various libraries including GDAL/OGR, Proj4, GEOS, JTS Topology Suite and GeoTools. Also tested application including UMN Mapserver, OSSIM, QGIS, PostGIS, GeoServer, uDig and GRASS. Among the several possibilities UMN Mapserver was selected for the project.
Beowulf Cluster:
Administer the research group cluster. The cluster consisted of 4-node, 8 processor, x86_64 architecture. Configuration included Fedora 6 setup on local OS drive, 4TB Raid 5 home storage mounted via NFS, NIS based login, GigE switch configuration for interconnect, DMZ for external network. Software application installed included OpenMPI, Matlab, Octave, Intel IPP, GNU Compiler Suite, Intel Compiler Suite and research group developed software porting and installation.
Prediction of Engineering Properties of Continuous Fiber Reinforced Composite Lamina:
This was my senior design project. The project involved developing a Finite Element (FE) model of a graphite fiber-reinforced polymer(GFRP) lamina. Used the model to predict the variations in the engineering properties of the lamina with the variation in fiber angle. Predictions were compared to analytical model of an angle-ply lamina and were found to be in close agreement.
python-gnome2-extras python-feedparser python-gdata python-webkit