Projects

Georgia Tech:

• • Multi-Robot SLAM ( Dr. Frank Dellaert ) -- Goal is to develop a method to efficiently perform SLAM across aerial and land robots. This is part of a bigger collaboration for the MAST project. The challenges are manifold as there is a limit on the communication bandwidth, the quality of the sensors used etc. The research is at a fundamental level and we are exploring various ways to tackle the distributed aspect of SLAM.

• High Fidelity Localization and Navigation ( Dr. Henrik Christensen ) -- Goal was to develop a high fidelity localization and navigation system using only laser scanners for the Boeing Paint hangar project. The system was demonstrated  at Boeing's Everett paint hangar. The videos are confidential so I am sharing the videos of the Segway using the same algorithm in the Georgia Tech Lab.

• EasySLAM ( Dr. Frank Dellaert ) -- As the name suggests the objective is to provide an easy way to do reliable SLAM in indoor environments. The project is divided into multiple phases with increasing levels of difficulty. The first stage is to use known markers followed by known palanar markers follwed by extracting planar regions automatically and so on. The end product will be an open source SLAM tool which will be available to download and will allow you to capture and interact with any kind of robot. ( There is a neat hack we use to circumvent the third party capture softwares! :)

• Visual Place Categorization ( Dr. Jim Rehg ) -- Goal was to understand the existing Visual Place categorization system and come up with improvements. Due to time constraints new suggestions were not implemented but this project provided me a good insight into various feature detectors especially CENTRIST which was developed by one of Jim's former students.

• Category Specific Super Resolution ( Dr. Charles Isbell ) -- The idea of super resolution is quite old. Single image super resolution is usually achieved by using Machine Learning techniques. The training image used for these are usually some randomly chosen examples. The project explored the idea of specifically selecting a category and using those images and see if that improves the quality of super resolution. We used Viola & Jones based face detector and parts of faces (eyes, nose, lips) detectors and created a hand-tailored training dataset and used it specifically on test images from these categories. The results were better than randomly trained model.[pdf]

IBM:

• Anomaly Detection for Surveillance ( Dr. Yun Zhai, Dr. Sharathchandra Pankanti ) -- The goal of the project was to come up with an online method to detect anomalies in surveillance data. The details of the work are show in the presentation:[ppt]

• Anomaly Detection for Rail Data ( Dr. Ying Li, Dr. Sharathchandra Pankanti ) -- The goal of the project was to come up with an online method to detect anomalies in railway data. I focussed mainly on anomaly detection in tie-plates and used texture features(Gabor) and a combination of kmeans and hierarchical clustering to find salient regions and do anomaly detection. The entire process was unsupervised and the method had to incrementally adapt to the data and change the definition of anomaly. Below is an image of the anomalies detected on a small dataset ( red labels ) alongside normal tie-plates. 

Sarnoff Corporation:

• Object Segmentation and Matching from aerial Lidar data ( Dr. Harpreet Sawhney, Dr. Bogdan Matei ) -- Goal is to develop a method to segment objects from aerial lidar data and identify the objects to assist in 3D modeling of cities. Few example classes include, Buildings, Trees, Vehicles, Traffic Signs, Lamps etc. The images and videos cannot be made public so have not been put on the site due to restrictions.

• Indoor Mapping and Retrieval ( Dr. Rakesh Teddy Kumar, Mr. Jayan Eledath ) -- There were two parts to the project. The initial part was to use the Pioneer robotic platform with a mounted Directed Perception head and Hokuyo Laser to build point clouds using Odometry and Iterative closest point algorithm. The second part involved segmenting out known objects from the point clouds. Chairs, Stairs and Doorways were few objects of interest. I used spin images to do this and the performance was quite good for chairs and stairs. The images and videos cannot be made public so have not been put on the website due to restrictions.

  • Automated Cinematography for Sports ( Mr. Jayan Eledath ) -- www.pajamachannels.com explains the objective of this project. It is a real time sports system which uses vision algorithms for automated coverage. It has been my dream project because of two reasons, I was part of the small team to ponder on the idea and prove the first prototype. It is a live system developed from scratch and deployed at multiple colleges. For a graduate student willing to spend the entire day at work, to see it going live is a dream come true. An experimental football output is also available. 

• License Plate Detection ( Dr. Manoj Aggrawal ) -- Goal of the project is to create a real time license plate detection system in an uncontrolled environment. The approach was based on Difference of Log Gabor responses. The main objective was to expedite the algorithm without losing out on accuracy. Integral images along with OpenMP based loop parallelization were used to achieve required performance.

• VideoHTML ( Mr. Arvind Lakshmikumar ) -- Goal is to create a robust tracking algorithm to handle pose and scale variations, occlusions and work in low resolution videos. Objects are tracked using covariance features and EM shift in a semi-automated fashion to create interactive videos. The viewers can interact with the video and get respective annotated information. The website can be accessed at www.stickypixel.com

• Fragment Tracker ( Mr. Arvind Lakshmikumar ) -- We combined best of object representation and object searching strategies in tracking to create a hybrid tracker. Object to be tracked is first broken down into multiple fragments in HSV space and each of the fragment is tracked using EM shift approach. The tracks of the fragments is combined to get the holistic track output.

• Hand Tracker ( Mr. Arvind Lakshmikumar )-- The objective was to track hand contour and recognize the gestures. This could be used as an alternative input to the pc, gesture recognition etc. The tracking was done in HSV space using covariance features, the edge operators are used to get the contour and points on the contour are tracked using condensation algorithm.

Work done in free time:

• Sports classification using Geometric properties-- This project is about classifying sports images using the invariant nature of cross-ratios under projective transformation. The algorithm works as follows: A histogram of cross-ratios computed based on the intersection of lines detected using Hough transform is used to form a feature vector of the image. A modified One-Against-All multi-class SVM classifier is used to classify the feature vector. Images of Tennis, Football, Basketball and Badminton were used for classification. This algorithm is published in Asian Conference on Computer Vision (ACCV)'07.

• Stroke Classification in Tennis -- Classifying the different tennis strokes is the prime goal of this project. We came up with an algorithm that uses the gradient information of the player's skeleton. The player is modeled using color histogram and tracked across the video using histogram back projection. An oriented histogram of the skeleton obtained in each frame forms the feature vector which is then sent to a trained SVM classifier to classify the strokes: Forehand, Backhand and No Shot. The work got published in International Conference on Image Analysis and Recognition (ICIAR) '07.

IIIT:

• Palmprint recognition ( Dr. Anoop M. Namboodiri ) -- Developed a real time biometric system which handles pose variations and deformations of the palm. Pose estimation and elastic matching techniques were used to handle the variations. Also, created a new palmprint dataset which contains the pose variations and deformations as there was none existing.

• Text based video search ( Dr. C. V. Jawahar ) -- Analyzed various text extraction methods for images and videos and designed a hybrid approach combining the advantages of bottom-up and top-down methodologies. Invariant features are computed for the extracted text regions and indexed for the search process. The novelty of the system lies in using the DTW based word-matching technique instead of extracting text for the retrieval process. The approach was tested on three languages, English, Hindi and Telugu.

• Face recognition ( Dr. P. J. Narayanan ) -- Studied and experimented with various face recognition methods to come up with an Illumination invariant recognition technique. Principal component analysis, Fischer linear discriminant, Neural networks, DCT were few of the methods tried. The concept of illumination cone was explored and combined with PCA to get higher accuracies though illumination invariance was not completely achieved. Also, created a dataset with controlled illumination conditions in the process. The report contains the work and the sample images of the dataset.[pdf][pps]

• Digital watermarking and Stegenography ( Dr. Jayanthi Sivaswamy ) -- The objective was to come up with a watermarking technique which has a duel role of helping in copyright protection as well as metadata storage. Explored existing watermarking techniques and adapted them to serve the purpose. %An application was also developed that allowed to embed and decode the watermark. It was tested  for medical and sports images. The system has the advantage of having the metadata in the image instead of an additional data source.[pps]

Image Compression ( Dr. V. Ch. Venkaiah ) -- Goal was to evaluate compression techniques and find the best method to encode vision datasets. Studied various compression techniques like JPEG 2000, Fractal compression, Differential RLE. The tradeoff between time and compression was the key to us. Also, evaluated PCA based compression which worked well for Face and Palm datasets.[pps]

Graphics toolkit ( Dr. P. J. Narayanan ) -- The project consisted of series of assignments involving 2D graphics, 3D graphics and finally writing own toolkit to draw lines, polygons. Polygon filling, clipping using Z-buffering were also part of the tasks.