Dhrumil Kotadia

Dhrumil Kotadia

dkotadia@wpi.edu

Projects and Research

Vision Based Collision Risk Estimation

Designed a semantic segmentation-based system utilizing Mask R-CNN, SORT for real-time tracking, and pedestrian velocity prediction to generate danger zones around pedestrians. The resulting system estimates collision risk at 10 FPS.

Technologies: Python, Pytorch, OpenCV, Scipy, Mask R-CNN, SORT

Click here for more info

Stereo Visual Odometry

Designed a scalable monocular visual odometry system for the estimation of the 6 DOF pose of the ego vehicle. Implemented ORB feature matching, depth estimation, triangulation and Bundle Adjustment, processing 12 fps on the KITTI dataset.

Technologies: C++, OpenCV, Eigen, Ceres

Click here for more info

3D Scene Reconstruction using Nerf

Designed and implemented a Neural Radiance Fields (NeRF) deep learning model to generate novel views of scenes from images, allowing the creation of 3D object GIFs with photorealistic rendering and seamless view synthesis.

Technologies: Python, Pytorch, OpenCV, Scipy

Click here for more info

Object Detection Enhanced Visual Odometry

Improved feature based monocular visual odometry system using dynamic object detection. ORB features from dynamic objects are filtered using YOLOv8 in the scene. Achieved 10% enhancement in the overall estimated trajectory.

Technologies: Python, OpenCV, Pytorch, Numpy, Scipy

Click here for more info

Monocular Visual SLAM

Developed a Visual SLAM system for autonomous vehicles. Leveraged feature-based methods (ORB), loop closure detection with Bag-of-Words and Bundle Adjustment, reducing trajectory drift by 20%.

Technologies: Python, OpenCV, DBoW, Numpy, Scipy

Click here for more info

Panorama Generation - Feature Based Image Stitching

Designed a module to create a panoramic image from multiple images(up to 6) using Harris Corner detection, Feature descriptor generation, Non Maximum Suppression, Feature matching using feature descriptors and RANSAC.

Technologies: Python, OpenCV, Numpy, Matplotlib

Click here for more info

Sparse 3D Reconstruction using Structure from Motion

Designed a Structure from Motion pipeline leveraging epipolar geometry constraints, RANSAC outlier rejection, triangulation, PnP, and Bundle Adjustment. Improved camera pose estimation accuracy by 85%, reducing reprojection error from 2720.6 to 381.09 pixels.

Technologies: Python, OpenCV, Numpy, Matplotlib

Click here for more info