🛠️ Self-Supervised Projects 🔧

Lidar Obstacle Detection with ROS

• Goal: use various algorithms on Point Cloud data such as Voxel Grid filtering, RANSAC segmentation, and Euclidean Clustering with KD-Tree to detect obstacles

• Pipeline

  

• Results The following animation shows the segmented point clouds - obstacles (in yellow) and road (in green)

  

Autonomous Maze Solver Robot

• Developed an autonomous maze-solving robot using TurtleBot3 and ROS
  
• Integrated RPLIDAR sensor for real-time obstacle detection and avoidance
  
• Implemented intelligent wall-following and search algorithms
  
• Leveraged TurtleBot3 for rapid prototyping and navigation strategy testing
  

(Video 1) (Video 2) (Video 3)

Real-Time Mask Detection using Convolutional Neural Networks (CNN)

• Developed a real-time facial mask detection system using a custom Convolutional Neural Network (CNN) model built with TensorFlow
  
• Designed and trained CNN architecture to accurately identify the presence of masks on faces in video streams
  
• Optimized model for real-time performance, achieving <100ms inference time per frame on webcam feeds
  
• Achieved 96% accuracy in detecting masked individuals in real-world conditions, ensuring reliable monitoring
  
• Demonstrated expertise in computer vision, deep learning, real-time analytics, and deploying AI solutions
  (Video) (Code)

🎓 Mentored Projects 🎓

VR-Based Smart Whack-A-Mole: A Benchmarking Platform for Target Prediction Algorithms

Aug’22 - Dec’22

Overview: Developed a VR-based Whack-A-Mole benchmarking platform in Unity, integrated with Oculus and Leap Motion IR for egocentric hand tracking.
Contribution

• Collected a large-scale egocentric hand-target dataset from 4 users (2,900+ trials at 100Hz), including 21-finger centroid trajectories across scaled inter-target distances (0.7–1.0).
• Implemented and benchmarked multiple target prediction algorithms: analytical velocity-based, DTW (91% test accuracy), and LSTM/RNN models (up to 90% real-time AI scores). *Designed standardized evaluation metrics: prediction speed distributions, multi-class accuracy, confidence scores, and trajectory point MAE (0.0018).
• Conducted robust real-time testing (100 trials/model at unseen scales 0.58–0.9), demonstrating LSTM models outperform DTW in efficiency and real-time performance.
• Created an interactive game environment for quantitative benchmarking of prediction algorithms, enabling intuitive evaluation of egocentric hand tracking models. (Paper)

Modelling and Control of a Quadrotor System

Final Year Project, IIEST Shibpur | August 2020 – May 2021
Supervisors: Dr. Ashoke Sutradhar, Professor, IIEST Shibpur

Overview: Design a dynamic simulation model for a stable flight control for a quadcopter in SIMULINK
Contribution

• Formulated the 6-DOF motion equations using the Newton-Euler method, modeling thrust, drag, gyroscopic effects, and actuator limitations.
• Designed a modular Simulink plant model with realistic constraints including motor saturation, thrust–weight ratio (>2:1), and body–inertial frame transformations.
• Designed and tuned cascaded PID controllers for altitude, attitude (roll/pitch), and yaw, achieving stable tracking under step, ramp, and multi-axis trajectory inputs.
• Simulated realistic flight behaviors including hovering, altitude ramps to 5m in <10s, 12° attitude maneuvers, and 3D inertial trajectories, validating controller robustness across coupled dynamics.
  (Report)

Stabilization of Single Link Manipulator

IIEST Shibpur | December 2019 – January 2020
Supervisor: Dr. Aparajita Sengupta, Professor, IIEST Shibpur

Overview: Developed a nonlinear dynamic model and stabilization controller for a single-link robotic manipulator using Lagrangian mechanics

• Derived the nonlinear dynamics using Lagrangian and Euler–Lagrange methods, obtaining the second-order system $\tau = ml^2\ddot\theta + mgl\sin\theta$

• Linearized the system around the unstable equilibrium at θ = 45°, producing a state-space model with an unstable pole at +2.081 rad/s
• Designed and tuned a PID controller (Kp=54.5, Ki=15.3, Kd=48.5, N=60) to stabilize the manipulator for target angles between 30°–90° from initial states up to 0°–90°, achieving settling times under 3 seconds
• Validated performance through multi-angle step responses (0→45°, 0→90°, 60→45°, 90→30°), demonstrating zero steady-state error and low overshoot beyond the linearization region. (Report)