Robust Detection and Tracking of Humans

The aim of the project is to make a humanoid robot which should be able to detect humans and then interact with them. The current implementation uses Isard and Blake's contour based condensation tracking. Detection is based only on single frames. Now I am working on robustly detecting humans using multi-modal techniques, adding motion cues like periodicity and aspect ratios, dress models etc.

 

Multimodal Localization of Bats

The aim of the project is to find 3-D positions of a flying bat and track it using multiple calibrated views, and noisy correspondences.  Additionally, there may also be data available from a complementary modality, such as acoustics, that can be used to provide a second estimate for the object position. In the present case, we take this modality to be acoustical emissions from the tracked object that are recorded at an array of microphones. We are able to achieve accurate reconstructions of the bat's flight path using the developed algorithms, which are capable of real-time execution.

 

Gait Based Human ID

 

The project targets to identify humans on the basis of their gait. There are two general approaches to gait recognition (and motion based recognition in general). The first class of approaches considers gait to be comprised of a sequence of poses of the person, and recognizes it by recognizing a sequence of static configurations. For example in Tsai:1994, Tsai et al. use one complete gait cycle as a model and match it with one cycle of an input trajectory for recognition. The second class of approaches considers the entire gait as a spatio-temporal pattern and attempts to characterize this pattern for recognition. The drawback of the former is that it fails to capture the temporal and spatial relationships among successive poses of a person. We believe that much information about motion patterns of the gait are encoded in these relationships, which is achieved in the latter type of approaches.

 

Real-time Color Calibration of Digital Cameras

Position: Research Intern

Manager: Anoop Gupta    Mentor: Ross Cutler

Research Lab: Collaborative and Multimedia Group, Microsoft Research

Period: May23rd, 2001 - Aug 30th, 2001

 

The aim of the project was to develop an application to do color calibration of digital cameras, i.e. make the output image from the camera "look good". The second part of the project was to make an application to propagate the color parameters from one camera to another which have some overlap using the color features of the overlapping region. The work was published as a technical sketch in CVPR, 2001. For details and results, Click Here

 

 

Audio-Video Correlation Based Speaker Detection

Position: Research Intern

Manager: Anoop Gupta    Mentor: Ross Cutler

Research Lab: Collaborative and Multimedia Group, Microsoft Research

Period: May23rd, 2001 - Aug 30th, 2001

 

The aim of the project was to be able to detect a speaker, using single camera and single microphone, based on the audio and video correlation between the speech and the lip movement. We re-implemented some existing work by John Hershey to see how well it performs and then worked on enhancing it to do better. For some preliminary results, Click Here

 

 

Real Time Computer Vision Applications

 

Position: Student

Supervisor: Prof. Subhashis Banerjee

Research Lab: Computer Vision Lab, Indian Institute of Technology Delhi, India

Period: August 1999 - Feb 2000

   

This was my B.Tech. project. The aim of the project was to develop real-time vision applications. Part of the work was an extension of the work I did during my internship with Prof. Shmuel Peleg in Hebrew University of Jerusalem, Israel (HUJI). Using the real-time optical flow based registration implementation I developed at HUJI, I implemented Motion Segmentation and Image Stabilization.

 

In the remaining 3 months of the project I implemented Omni-Cam, similar to Shree K Nayar's work. Further, I combined this with my experience of Panoramas I gained at HUJI to implement "Omni-Panoramas". For details and results, Click Here.

 

 

Real Time Stereo Panoramas

 

Position: Research Intern

Supervisor: Prof. Shmuel Peleg

Research Lab: Computer Vision Lab, Hebrew University of Jerusalem, Israel

Period : May 2000 - August 2000

 

In this project I implemented from scratch, software to make Panoramas and Stereo Panoramas from a single hand held camera in Real-Time. The challenge in the project was to add blending to the incoming panorama strips, to normalize the intensity variations due to flicker of light sources. I successfully implemented blending using Gaussian pyraminds. For details and results, Click Here.

 

Fuzzy NL Query Based Search Engine

Position: Student

Supervisor: Prof. Saroj Kaushik

Research Lab: Artificial Intelligence Lab, Indian Institute of Technology, Delhi, India

Period: Jan 1999 - May 1999

 

In this project I implemented a search engine that could take as inputs, Fuzzy Natural Language Queries. The search engine worked on top of Postgres Database. The search engine was interactive and took user's help to disambiguate input queries if required and also learned in the process. Thus it was  a self learning system, and also had some self correcting features to make it robust to incorrect learning. For details of theory & implementation, Click Here.

 

Online Research Database for CS Deparment, IIT-D

Position: Winter Intern

Supervisor: Prof. Subhashis Banerjee

Research Lab: Intel Technology Lab, Indian Institute of Technology, Delhi, India

Period: Dec 1999 - Jan 2000

 

In this project I created an Online Research Database and a search engine on top of it for the computer Science Department of Indian Institute of Technology, Delhi, India. The database contains all the projects going on in the CS Department of IIT-D and all publications. The search engine can be used to search for projects, ongoing research and publications of the Department. This database can be updated and searched on the web Here.