Sadegh‑Kalami

Drone Localization:

The objective of this project is to develop a compact system utilizing embedded systems such as NVIDIA Jetson Orin, NVIDIA Jetson Nano, and etc, equipped with a hybrid model that integrates deep learning techniques to enable drones to navigate autonomously in the absence of Global Positioning System (GPS) signals.

Note: Code cannot be shared due to project funding and confidentiality constraints.

Demo:

Visual-Inertial Odometry:

After extensively researching various Visual Inertial Navigation Systems (VINS) like Visn-Fusion, Kimera, ORBSLAM3, Pyslam, among others, I carefully evaluated their capabilities and ultimately selected ORBSLAM3 for its impressive combination of robustness and accuracy.

To implement ORBSLAM3, I successfully deployed it on both Jetson Nano and Jetson Orin platforms. This involved adapting the system to run efficiently on these hardware configurations, ensuring optimal performance.

For rigorous testing and evaluation, I utilized diverse datasets representing both aerial and ground scenarios. These datasets included FGI Masala Stereo-Visual-Inertial Dataset 2021, EuRoC dataset, data collected using a Our Phantom 4 Pro, and a custom dataset generated with our in-house data collection system.

To showcase the effectiveness of my implementation, I've prepared a demonstration that provides a clear example of the successful application of ORBSLAM3 in various real-world scenarios. This demonstration serves as a testament to the system's performance and its adaptability across different datasets and platforms.

IEEE Transactions on Robotics 2021

ORB-SLAM3: Paper | Code

Demo:

Optical Flow:

Implementation of Optical Flow on our data captured by IMX219-160 camera using ORB methood to estimate displacement of robot which has been used in Robot Localization project.

Demo: The video playback speed has been doubled

IMU-Fusion:

Implementation of 9 DoF Inertial Measurement Unit (IMU) sensor fusion using Extended Kalman Filter (EKF) with Magnetometer and Accelerometer calibration. Usable for different types of IMU sensors such as MPUXXX, ICM20948 and etc.

Code

Demo:

3D Reconstruction Application:

As a funded project by IRIB we've developed a state-of-the-art 3D reconstruction application that operates in near real-time, utilizing image and video feeds to dynamically create 3D models of diverse subjects such as environments, objects, bodies, and faces. Employing cutting-edge technologies like Nerf, Gaussing Splatting, ECON, HRN, Meshroom, Avatar. To improve user experience, we've designed a user-friendly graphical interface based on Qt5.

Our project has garnered recognition and has been nominated for the Distinguished Service award, showcasing its excellence and impact in the field.

Note: Code cannot be shared due to project funding and confidentiality constraints.

Demo:

Smart Refree:

As a funded project by Tradional Sports Federation of Iran we've developed an intelligent referee AI system to count and assess the correctness of athletes' movements with a high degree of accuracy. To achieve this, we employ pose estimation deep learning models that accurately identifies key points on the athlete's body. This information is then subjected to geometry-based calculations.

The culmination of this process is a comprehensive display of results on the stadium scoreboard in Persian. By employing our AI-based system, we aim to eradicate the potential for human error and biases in the scoring of athletes, ushering in a new era of fairness and accuracy in sports evaluation.

Note: Code cannot be shared due to project funding and confidentiality constraints.

Demo: