Ajith Kumar Jayamoorthy

Deep Learning Intern - ADAS, Veoneer US Inc.

Real-time Radar-based Obstacle Detection using Deep Learning

• Improved near-distance obstacle detection accuracy by 2%, using modified loss function.
• Assisted in evaluating radar tracker data using RMSE and track association model
• Alternate approaches for anchor points for detection boundaries based on limitations of radar data

Image credits: Veoneer LLC

Machine Learning Engineer, Veoneer India Pvt. Ltd.

Interior Cabin Sensing for car's occupant detection and classification

• Constructed FCNN training framework using Python, achieved 99.5% accuracy in MATLAB evaluations.
• Supported C++ implementation for real-time testing, accomplishing 97.5% accuracy in multi-occupant scenarios
• Performed feature Analysis and derived feature development
• Developed training material and served as a part-trainer for an in-office machine learning training program.

Video credits: Veoneer LLC

Machine Learning Intern, Veoneer India Pvt. Ltd.

Interior Cabin Sensing for car's occupant detection and classification

• Extracted significant features from radar data and optimized Logistic Regression model training
• Verified model generalization, achieving 97.8% accuracy with minimal test data

Image credits: Veoneer LLC

Filter Banks and Boundary Detection

A simplified version of pb, known as pb-lite is developed for boundary detection. In this method the boundary is evaluated by examining brightness, color, and texture information across multiple scales. The output of this algorithm will be a per-pixel probability of boundary.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Auto-Calibration of Camera

This project is the implementation of Camera Calibration technique from a paper by Zhengyou Zhang of Microsoft. 13 images of Checkerboard pattern taken from Google Pixel XL phone is used for this calibration. The approximate intrinsic and extrinsic parameters of the camera are calculate and the Non-linear Geometric Error Minimization is performed to optimize these parameters. After optimization, the new parameters are used and the points are re-projected on the warped image

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Semantic Mapping of Point Clouds and Point-to-point ICP

This project comprises of implementation of two concepts. First, semantically segmented images were used to paint the corresponding point clouds. Second, the painted point clouds from various consecutive timestamps were combined using the Iterative Closest Point Algorithm.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Panorama Stitching

This is a traditional approach. Initially the corners are detected for each image that requires to be stitched. Then adaptive non-maximal suppression method is used to find local maxima of corners. Features are extracted from each image and then matched. Matched features clubbed with random sample consensus, and the outliers are removed using RANSAC. Homography is estimated and then the images are blended together.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Face Swap using Triangulation and TPS

Two traditional methods are used in this project for face warping, namely, triangulation and thin plate spline (TPS). For triangulation, we used Delaunay Triangulation method, which is the dual of Voronoi triangulation.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Structure From Motion - 3D Reconstruction

The main objective of the project is to generate a 3D model of a particular scene from multiple 2D images captured from different perspectives. Here, 5 images have been given to start with, a text file describing the 2D image point correspondences between all possible image pairs and the calibration matrix of the camera used for capturing the images.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Neural Radiance Field - 3D Reconstruction

The Neural Radiance Field or simply NeRF is the state-of-the-art deep learning method that generates a complex novel view of a scene by optimizing the underlying volumetric scene function using a sparse set of input images. The data for NeRF is given from the original authors. The input data consist of three files, namely, train, test, and validation with images consisting of a lego model of an object. Each Camera parameter for each image has been further given by .json files to be used for the training purpose.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Visual Inertial Odometry

This project is a implementation of a filter-based stereo visual inertial odometry that uses the Multi-State Constraint Kalman Filter (MSCKF). The mathematical concepts of a stereo-MSCKF are implemented within a frame-work of given starter-code.

credits: WPI RBE549 - CV by Prof. Nitin J. Sanket

Collision avoidance using various types of Velocity Obstacle method

The following project is the implementation of collision avoidance in Dynamic environment using the global-local planner integration. The python implementation of global and local planner has been done. For global planner the RRT* algorithm has been implemented. In case of the local planner, three variants of velocity obstacle has been implemented,namely, Standard Velocity Obstacle (VO), Reciprocal Velocity Obstacle (RVO) and Hybrid Velocity Obstacle (HVO).

credits: WPI RBE550 - Motion Planning by Prof. Zhi Li

BFS,DFS,Dijkstra and A*

The basic search algorithms have been implemented using a custom environment. The Basic search algorithms comprises of the following:

• Breadth First Search Algorithm (BFS)
• Depth First Search Algorithm (DFS)
• Dijkstra Algorithm
• A* Algorithm

credits: WPI RBE550 - Motion Planning by Prof. Zhi Li

PRM,RRT and RRT*

The standard search algorithms such as probabilistic road maps and rapidly-exploring random tree are implemented here. The Standard search algorithms comprises of the following:

• Probabilistic Road Map (PRM)
• Uniform sampling
• Random Sampling
• Gaussian Sampling
• Bridge Sampling
• Rapidly-exploring Random Tree (RRT)
• Rapidly-exploring Random Tree* (RRT*)

credits: WPI RBE550 - Motion Planning by Prof. Zhi Li

Informed-RRT* and D*

As the part of the advanced algorithms Informed-RRT* and D* have been implemented using custom enviroments.

credits: WPI RBE550 - Motion Planning by Prof. Zhi Li