• What are the efficient motion strategies for robot to collect data from sensor nodes?
• How can we further reduce the energy consumption of sensor nodes by searching a good location with robot to start download?
• How can we adapt the beacon schedule of sensor nodes according to the uncertainty in the motion of robots?

Please see Data Mules page for more information. Related publications [J3][C2].

• One such scenario is if the user which requires connectivity is a mobile entity (e.g. human, robot). If the area that target moves on is large compared to the wireless range of routers, robotic routers (mobile robots with wireless communication capabilities) can create an adaptive wireless network and provide communication services for mobile users on-demand. We consider the motion strategies of robotic routers to provide a continuous communication for user while minimizing the number of robotic routers.

• Another scenario when mobile robots can be helpful is when a communication bridge between two locations is needed and an underlying communication infrastructure does not exist. We consider the bi-criteria optimization problem where the objectives are minimizing the number of mobile hubs while minimizing distance traveled by the hubs.

Please see Robotic Routers page for more information. Related publications [J2][C1][C3].

Please see Sensor Placement page for more information. Related publications [J1][C4].

This project was a joint project with Nikhil Karnad for the Computer Vision course. An iRobot Create robot is programmed to go straight and capture frames along the way. The pixel points between two consecutive frames are matched using David Lowe's SIFT feature matching program. The best 10 matched points are selected with RANSAC and fed to 8-point algorithm. The four-way ambiguity is removed by using the odometry information. Kalman filter is applied to odometry measurements and movement estimation calculated from 8-point algorithm.

In this project, we assume that robot has only an IMU (Inertial Measurement Unit-Gyro+Accelerometer) and GPS to determine its position in world coordinates. Due to the uncertainty in the measurements (about 3m uncertainty in GPS) measurement from these sensors are fused to get a better estimate of the position. Assuming the bearing information is accurate (we found empirically that relying on only gyro provides an accurate bearing estimate +/- 2 degreees), navigation system is modeled as a linear system and Gaussian noise assumption is made. The performances of Kalman Filter and Particle Filter are compared. Although Kalman Filter is the optimal linear estimator under these assumptions, the performance of Particle was very close to Kalman Filter which can be used for non-linear systems and with any noise distribution.

The point matches between two scenes are generated manually. Computing the fundamental matrix directly behaves rather poorly hence we normalized the point matches. 8-point algorithm is implemented to calculate the fundamental matrix. Top two figures show the selected points in the first image and their corresponding epipolar lines in the second image. Using calibration matrices the essential matrix is calculated. Four-way ambiguity is resolved using ubiquitous SVD (Singular Value Decomposition). Bottom left figure shows the select polygons in the original image and bottom right figure shows the reconstructed polygons,

The goal of this exercise was to show how to reconstruct a surface given multiple photographs of the same scene under several known lighting conditions. I took advantage of Peter Belhumeur’s face image database. My procedure was to extract normal/albedo at each pixel and reconstruct the surface from normals. Left image shows the original image and right image shows the 3D reconstruction of the person's face.

libcwiimote is used for communication with wii controller and libcreate is used (which is created by myself) to control the robot. Extern is a very useful keyword to combine C and C++ codes. Controller works like a joystick which is rooted at the bottom of wii controller. Check out the video to see how it works.

RSSI (Radio Signal Strength Indicator) measurements from various locations and orientations are collected (50 measurement from each location and orientation). A pdf is extracted from this experiment, i.e. p(x|z) where x is the state vector (location and orientation) and z is the observation (observed rssi value). Particle filter algorithm from Ioannis Rekleitis tutorial is implemented in Matlab. Partticles are moved according the odometry values and the weights of the particles are updated according to the pdf. In the video, red dash line shows the grand truth and black arrows show the estimations.

Trajectory of the puck is calculated using dynamics properties of Air Hockey table and initial direction of the puck after it is hit by opponent's mullet (user controlled). The final location of the puck in the computer controlled mullet's side and arrival time is calculated according to this trajectory. A desired trajectory is found for the puck after hit by computer controlled mullet such as target location of the puck is left of goal, puck should hit to boundary twice to confuse opponent, etc. To satisfy this trajectory a final configuration for computer controlled mullet is found, such as location, velocity and time to hit. A third order polynomial is used to calculate the forces to apply to reach the final configuration from the current configuration.

This system was implemented for a visit of Capital District Middle School students.

Implemented a GPS software using ESRI’s ArcPad on a PDA running Windows ME and implemented a server program using ESRI’s MOJava to update geographic information (second best senior project award). For more information please visit our group web site

• Developed a feature extraction scheme using Haar wavelets for face detection
• Implemented eigenface algorithm for face recognition

This work was done under the supervision of Prof. Fatos Yarman Vural

During the depth scan of faces, hairy parts (mustache, eyebrow, etc) of the face can not be detected which requires a preprocessing step before feature extraction. Interpolation and morphological operators does not work well since they cause to loose the smoothness of the face which affects the performance of our face recognition algorithm. I have used Snake algorithm to recover missing parts while preserving the curvature and smoothness of the face. This work was done under the supervision of Prof. Charles Dyer