Eric Perko's Homepage

Course Projects

A collection of miscellaneous course projects that were geared towards applications in robotics.

Real-time Object Detection for IGVC for EECS 600 Machine Learning, Fall 2009

Abstract: Detection of arbitrary objects in images is a difficult task, made even more so by the real-time detection requirements of a robot designed to compete in the IGVC. For a robot in this competition, objects need to be classified in real-time so that obstacles can be found and avoided before a collision occurs. In order to evaluate a possible solution to this problem, I evaluated the performance of a Viola- Jones object detector implemented in OpenCV and an extension of that framework to the Multiple Instance Learning setting. I hypothesized that the Viola-Jones detector would achieve moderately good detection rates on my sample data and that the MIL variant would improve these rates. In order to test this, I trained and evaluated these algorithms on logged data from IGVC '09 and attempted to detect orange barrels, A-frames and white lines painted on grass. From these tests, I found that the performance of the Viola-Jones detector was actually very poor on this data set, most likely due to the feature representation used.

EKF SLAM Implementation and Estimation Evaluation for EECS 491 Artificial Intelligence, Spring 2009

Abstract: I implemented an algorithm to solve the online simultaneous localization and mapping (SLAM) problem using an extended Kalman filter (EKF). SLAM is an interesting problem in robotics that is made up of two much simpler problems - if you know where all of the landmarks are, you can decrease error in the robot pose odometry and if you know the robot pose, you can reduce the error in landmark positions from sensor data. The difficulty arises when neither the landmark positions nor the pose are known exactly and you wish to get both a good estimate of the current pose and the locations of all the landmarks at the same time – hence the name simultaneous localization and mapping. In order to evaluate the performance of this algorithm, I varied the total number of landmarks, the noisiness of the simulated sensor data and the number of times that the simulated robot looped through the map and compared the estimated locations of all landmarks with their actual locations. I observed that EKF SLAM is very sensitive to variance in the sensor data and that the number of loop iterations does not influence the time taken to converge to landmark errors with low errors. The key was the number of landmark re-observations – two re-observations seemed to be the magic number for the parameters that I tested.