Projects

Implemented a visual odometry pipeline to estimate autonomous vehicle trajectory using monocular camera images from the CARLA simulator.

Utilized ORB (Oriented FAST and Rotated BRIEF) algorithm for efficient feature detection and extraction across sequential camera frames.

Developed a brute force matching algorithm with L2 norm to robustly match ORB features between image pairs.

Optimized feature matching by implementing distance-based filtering to remove outliers, improving trajectory estimation accuracy by 15%.

Estimated 3D camera motion between frames using techniques like Perspective-n-Point (PnP) and Essential Matrix Decomposition to reconstruct the vehicle's trajectory.

Implemented sensor fusion using proprioceptive IMU sensor data and exteroceptive LiDAR and GPS data from CARLA simulator, applying Error-State Extended Kalman Filter to perform state estimation and localize ego vehicle. 

Altered transformations between LiDAR and IMU sensor frames to miscalibrate sensors, before adjusting filter parameters to output correct vehicle pose estimates while compensating for resulting noise variance. 

Tested Kalman Filter implementation by comparing estimated vehicle position to ground truth position, achieving 83% accuracy.

Utilized Canny Edge Detection and Hough Transform computer vision algorithms with OpenCV to perform lane detection through semantic segmentation on JPEG image and 31.9MB video.

Converted image to grayscale, applying Gaussian Blur OpenCV algorithm to filter image noise, modifying resolution by 33.2%.

Highlighted regions of high pixel gradience using Canny Edge Detection, masking 67.1% of pixels to isolate region of interest.

Developed autonomous ROS2 robot in C++ and Python that employs 3D pose estimation to detect, track, and follow a dynamic tennis ball in real-time while avoiding obstacles using A* path planning.

Performed SLAM with ROS2 slam_toolbox to generate 2D occupancy grid map of Gazebo environment using Nav2 for Adaptive Monte Carlo Localization in RVIZ.

Engineered object detection algorithm by applying HSV thresholding on RGB camera images before utilizing OpenCV SimpleBlobDetecter to detect tennis balls with 96% success.

Designed 3D pose estimation algorithm using trigonometry and intrinsic camera parameters such as focal length and field-of-view to estimate tennis ball’s 3D pose from 2D images and localize it in under 125ms.

Leveraged Nav2 to generate robust global costmap from 2D occupancy grid and plan optimal obstacle avoidance path using A* path planning.

Programmed object detection and tracking algorithm with Python, OpenCV, and YOLOv4 machine learning model with PyTorch to detect vehicles and track their trajectories.

Utilized OpenCV and NumPy to read in frames from 8.2MB video, generate green bounding boxes for each detected car, and assign unique tracking IDs to each car.

Simultaneously detected and tracked the motion of up to 14 vehicles per frame, with an accuracy of 86%. 

Designed Canny Edge Detector from scratch to segment lane lines and traffic signs in real-time.

Applied gaussian filter to the image to reduce noise and smooth out unnecessary details, leveraging  Sobel operator to calculate the intensity gradient in both the horizontal and vertical directions.

Developed an edge detection pipeline incorporating non-maximum suppression, double-thresholding, and edge tracking through hysteresis, using empirically tuned thresholds to minimize noise and enhance prominent edges.

Automated a 4-wheel mobile robot using C++, to perform localization, using an HC-SR04 ultrasonic sensor and gyroscope to navigate through a room detecting and taking orders from students before delivering food to them at their original locations.

Performed path planning by establishing robust and versatile data structures to store robot route instructions and order data, optimizing iterative algorithms to read-in stored data and generate optimal robot trajectory.

Accurately delivered 92% of orders within a 50cm2 margin of error from initial order location, to students in 17 locations.

Engineered an Arduino-based RC car with infrared and smartphone remote control using HC-05 Bluetooth module.

Programmed in C++, utilizing HC-SR04 ultrasonic sensor and SG90 servo motor for path planning and obstacle avoidance.