Jagennath Hari's Projects
This projects proposes a novel temporal attention based neural network architecture for computing visual odometry using the sequence of images.
LQR and iLQR controllers for a 2D quadrotor.
ConveyorVision is an innovative real-time system designed to automate the counting and tracking of cement bags on conveyor belts. Utilizing cutting-edge deep learning techniques like YOLOv8 for object detection and Byte tracker for precise tracking, ConveyorVision accurately monitors cement bags as they traverse the conveyor belt.
Harness the power of GPU acceleration for fusing visual odometry and IMU data with an advanced Unscented Kalman Filter (UKF) implementation. Developed in C++ and utilizing CUDA, cuBLAS, and cuSOLVER, this system offers unparalleled real-time performance in state and covariance estimation for robotics and autonomous system applications.
DepthStream Accelerator: A TensorRT-optimized monocular depth estimation tool with ROS2 integration for C++. It offers high-speed, accurate depth perception, perfect for real-time applications in robotics, autonomous vehicles, and interactive 3D environments.
A pdf tutorial for RTAB-Map
Dive into cutting-edge FusionSLAM, where SuperPoint, SuperGlue, Neural Depth Estimation, and Instant-NGP converge, elevating Monocular SLAM to unparalleled precision and performance. Redefining mapping, localization, and reconstruction in a single camera setup.
Implementation of inverted pendulum controller using Q-learning.
Implementation of different Kalman Filters to estimate the state of the quadrotor using optical flow and IMU.
[CVPR'24 Highlight] Gaussian Splatting SLAM
Graph based SLAM for multiple cameras using SuperPoint feature detector
Visualization of path planning algorithms(global planners).
Implementation of Extended Kalman Filter for quadrotor state estimation by fusing IMU and Vicon data
Computed Visual Odometry using corner extraction from April Tags and optical flow using ORB
Implementation of Unscented Kalman Filter to estimate state of quadrotor using optical flow and IMU.
The main goal of this project is to come up with an architecture having the highest test accuracy on the CIFAR-10 image classification dataset, under the constraint that model has no more than 5 million parameters.
RGBD-3DGS-SLAM is a monocular SLAM system leveraging 3D Gaussian Splatting (3DGS) for accurate point cloud and visual odometry estimation. By integrating neural networks, it estimates depth and camera intrinsics from RGB images alone, with optional support for additional camera information and depth maps.
A ROS package for Semantic Structure From Motion(SFM) using RGB Image, Depth Map, Camera calibration, Odometry and Semantic Segmentation Network.
TreeScan3D is an innovative project that leverages the power of 3D computer vision and multi-threaded processing for efficient tree detection. Built using C++ and integrated with the Robot Operating System (ROS), this tool excels at accurately identifying trees in complex environments.
Universal Monocular Metric Depth Estimation