This project implements a Perception module for ACME Robotics using high-quality software engineering practices such as the Agile Iterative Process, OOP, Pair Programming and Test Driven Development. We achieved near robust human detection using Histogram of Oriented Gradients (HOG) feature descriptor combined with a Support Vector Machine Model.
This project implements Stereo Rectification using Epipolar Geometry, SIFT, and Depth and Camera Pose estimation techniques to estimate the depth of the given 3D scene from two different perspectives. The Fundamental and Essential Matrix estimation is carried out using the 8-point algorithm and the RANSAC algorithm.
This is an end-to-end vision pipeline to swap faces in videos and images using warping methods Delaunay Triangulation and Thin Plate Spline. Face Landmarks are obtained by dlib then getTraingleList() and subdiv2D() are used to implement Delaunay method to swap. Thin Plate Spline model parameters are obtained to map Source to Destination face, then all Source pixels to Destination face pixels are warped.
Structure from Motion
This project presents our pipeline for
recreating a 3-D scene using Structure from Motion. Reconstructed a 3 dimensional scene
with 2D stereo images from a monocular camera captured from
different views while estimating camera poses along the way. The pipeline consists of Feature Matching, RANSAC Based Outlier feature rejection
and Estimation of Fundamental Matrix, Estimation of Essential matrix from F matrix, Camera Pose Estimation and Refinement, Check for Cheirality Condition using Triangulation, Linear and Nonlinear Perspective-n-point estimation, Bundle Adjustment to achieve the results.
recreating a 3-D scene using Structure from Motion. Reconstructed a 3 dimensional scene
with 2D stereo images from a monocular camera captured from
different views while estimating camera poses along the way. The pipeline consists of Feature Matching, RANSAC Based Outlier feature rejection
and Estimation of Fundamental Matrix, Estimation of Essential matrix from F matrix, Camera Pose Estimation and Refinement, Check for Cheirality Condition using Triangulation, Linear and Nonlinear Perspective-n-point estimation, Bundle Adjustment to achieve the results.
For Lane Detection, Image denoising is done using Gaussian Blur. Homography estimation is used to obtain the birds eye view of the road. HSV and HSL masking is used to obtain lane candidates and the plotted histogram of these masks provides maximum number of lane candidates to determine the lane eventually obtaining the turn and Radius of Curvature.
This project implements an Optical flow pipeline which shows the Motion Vector Field (top-right) with spaced red arrows showing the direction of motion features, the Dense Optical flow (bottom-left) of the approaching vehicles and background subtraction (bottom-right) using optical flow.
In this project I implemented state-of-the-art PB (Probability of Boundary) Lite Algorithm by generating DoG, LM, and Gabor filter banks, developing Texton, Brightness, and Color maps and then computing Texture, Brightness, and Color Gradients.
