Md Ashiqul Alam Khan
Forskning
Benchmarking of Underwater Image Dataset
Worked on a project centered on underwater image datasets with temporal variations. Focused on refining 3D image alignment algorithms for improved accuracy in challenging conditions. Also evaluating state of art image matching algorithms such as Super GLue, LightGlue, Loftr, Ecotr, DKM, and RoMA etc. Responsibilities include algorithm implementation, parameter tuning, and rigorous testing using our underwater image dataset. Main objective is to gain a comprehensive understanding of the distinctive strengths and limitations of each algorithm, particularly within the realm of underwater image analysis. Notably, the dataset captures temporal dynamics, offering a valuable perspective on how these algorithms perform under changing conditions.
Underwater Machine Vision for Long Term Operation of Robotic Platforms
Motivation for this project can be divided into two challenges
- Geo localization and temporal change detection of an underwater location or structure
- Visual aided navigation and localization of underwater systems
Qualitative analysis of Underwater Color correction Alogirthms
- Implemention underwater image enhancement algorithms
- Qualitative benchmarking of Color correction algorithms on key point matching
Developing an Autonomous Boat to follow predefined GPS locations, Guerledan Robotics
challenge
The Universite de Toulon and Ensta Bretagne organized an autonomous boat competition in Guerlédan Lake, France. Participants were tasked with constructing a boat that could follow a pre-determined path marked by GPS points in the shortest time possible. Our team emerged as the victors, securing first place in the competition. Main tasks of this project: 1) IMU and magnetometer calibration. 2) Vessel Motion control (Speed and Heading) based on GPS, IMU, Gyroscope and Magnetometer data. 3) Propeller voltage control based on sensor data. 4) Remote controller setup for manual boat operation.
Plant Classification Based on Leaf Recognition using Multilevel Residual Network and
Progressive Neural Architecture Search
In this study, a novel plant classification solution was proposed based on leaf recognition. The approach combined the strengths of Progressive Neural Architecture Search (PNAS) in deep feature extraction and architecture modularity with the simplicity of the ResNet 50 model. This resulted in a more effective plant classification solution compared to existing methods.