Hello I'm
Kavya Devadiga
I am a Robotics Masters' student at the University of Maryland College Park, USA. I like smart autonomous systems. I love to design systems that would automate the dull, dangerous, and mundane tasks and make a difference in human lives for betterment.
Projects
Hand-written Digit and Transfer Learning
Implemented classifiers for handwritten text recognition (OCR) and for the MNIST Monkey dataset using transfer learning. Different models built for ORC were SVM, Logistic Regression, and Deep Learning Networks. The Monkey MNIST dataset was classified using a ResNet50-based transfer learning network and results were compared with a simple CNN network classifier.
Face, Pose and Illumination Classifier
Implemeted label classification for three datasets: Face, Pose and Illumination. Bayes Classifier, KNN Classier, Kerner SVM and Boosted SVM algorithms are written from scratch in Python. The results were evaluated and compared to understand the effectr of dimentionality reduction techniques including PCA, LDA and MDA validation using K-fold cross validation.
AR Tag Detection, image superimpose and 3D projection
Detecting AR tag from a video and decoding its value by implementation of corner and edge detection from scratch. After detection, superimposing a png image on the tage. Next, a 3D cube is superimposed on the corners of the detected AT Tag.
Jackal Swarm Robot
A ROS based system that will deploy a swarm of 20 robots in a simulated gazebo environment and help rescue an object to the defined end goal. The swam consists of Jackal unmanned ground vehicles. Robots spawn at a pre-set start position and then navigate towards the object to be rescued. Once they reach the object contact point the agents will the object synchronously till the destination is reached.
Food-E : Diner robot
Modelling and simulation of an autonomous food serving robot in a diner. Using CoppeliaSim developed a 6 degree of freedom UR5 robotic arm fixed on a 2 degree of freedom four-wheeled moving cart model. Followed by the implementation of a four-wheeled differential drive, and forward-inverse kinematics and program the model to simulate the task for autonomously navigating and completing a pre-defined job.
CNN based Real-time Helmet Detection
Automated helmet detection using video surveillance and generating a database with license plates of offenders. Included dataset creation for helmets by web scraping and training a CNN module using TensorFlow, OpenCV based machine learning in python. Followed by KNN based text recognition for extracting the license plate data.
Curve fitting
Curve fitting using Least square, total least square and RANSAC from scratch.
Point robot path planning
Path planning simulation for a point robot to navigate from source to destination using Dijkstra algorithm.
Lane detection
Straight and curved lane detection and displaying next action command based on computed results.
Car CAD modelling and simulation
Modelled a car model in Solidworks. Visualized LIDAR points in RViz. Simulated PID controlled navigation in Gazebo using teleop control as well as publisher-subscriber in ROS. Understood basics of CAD modelling, URDF, XACRO, ROS-config files, ROS controllers and gazebo plugins, and adding sensors to CAD model.
Dual Pendulum LQR-LQG Controller
Worked on design and analysis of an LQR and LQG controller or pendulum attached to a frictionless moving crane. Using Control system concepts and Matlab.
Arm path planner
A 6DOF UR5 robotic arm path planner. Consists of a C++ based IK solver and planner module with output simulated in a virtual environment. The input to the system is the desired goal cartesian coordinates. This input is applied to a 6 DOF robotic arm with set initial points and known configuration. Based on this data, the IK solver calculates the configuration of individual arm joints.
DVR - Autonomous Robot
An autonomus - omni directional robot navigation through a controlled pre-described path using distance vector routing (for navigation) and PID algorithm (for line following). The robot has three pneumatically controlled arms to launch shuttles in air. The code is uploaded on Teensy development board (ARM Cortex-M7 at 600 MHz processor). (Project code was developed by a sub-team of 4)
Vader - Pole climbing robot
An Arduino based robot with power screw based pole climbing capabilities. The robot was controlled using a PS2-controller. It also has a vertical gimble with a high power BLDC motor to drive the ecobot (Project was developed by a team of 7)
PID line follower ecobot
An Arduino based line follower robot with single servo-controlled actuation. The robot has no drive of its own and relied on a second robot to power it without contact using wind thrust. It successfully maneuvered through a path consisting of turns, slopes and a very low contrast path-following line. (Project code was developed by a sub-team of 3)
Publication
Real Time Automatic Helmet Detection of Bike Riders>
International Journal for Innovative Research in Science & Technology | Volume 4, Issue 11, April 2018. ยท Mar 31, 2018
A CNN-based helmet detection system for real-time helmet detection. Using video surveillance of the street, the proposed approach detects if the bike rider is wearing a helmet automatically without manual intervention. If a bike rider is detected not wearing a helmet, the number plate of the vehicle read and noted. A database will be generated with records to identify every offender accurately by his bike registration number
Technology Stack
