Funded Projects
Some of the most exciting projects that I've worked on. Feel free to ask for more info.
PIRvision Occupancy Sensor
Role: Project Investigator
Ground breaking Passive Infra-Red (PIR) based Stationary Occupancy Sensor. With the help of patented solid-state IR shutter, on-board Machine Learning and low-power Edge AI device, this sensor helps reduce energy costs by 57% via accurate stationary occupancy sensing.
Funding Agencies: NSF, ARPA-E​
​Demos and Datasheet: pirvision.com
Simulation for a Mems-Based CTRNN Ultra-Low Power Implementation of Human Activity Recognition
Role: Primary Researcher
I developed a simulation model for a novel computing approach that uses Microelectromechanical Systems (MEMS) sensors for simultaneous acceleration sensing and computing. This innovative approach is aimed to eliminate the traditional computing elements such as the processor and RAM to reduce power consumption, size, and cost in wearable electronics and other resource-constrained devices for human activity recognition.
Funding Agencies: NSF
​Publications: [7], [8], [9]
GitHub Repo: github.com/em22ad/MEMS_CTRNN_learning_framework
Networked Smart Sensors Meet Bayesian Fusion
Role: Primary Researcher
This work presents a novel Bayes Filter-based occupancy detection framework that achieves superior accuracy without relying on historical sensor data. By modeling indoor occupancy as a Markov Decision Process and combining LSTM-based classification with Bayesian filtering, our system achieves 92.04% detection accuracy - outperforming existing methods by up to 15.52%. With an execution time of just 59ms per estimate and validated through extensive 30-day testing, this work demonstrates a computationally efficient and practically deployable solution for smart building applications, even in challenging environments with infrared noise and shielding effects.
Funding Agencies: ARPA-E
GitHub Repo: github.com/em22ad/Occupancy_Detection_Multisensor_Fusion
Drone based Hazard Level Estimation in Adversarial Environment
Role: Primary Researcher
Multiple robots are assigned a navigation task in uncertain conditions, and they may fail, potentially providing valuable information to mission planners. The approach involves modeling each robot's mission as MDP in a sequential deployment scenario. By incorporating mortality and communication imperfections, a Bayesian Estimator is constructed to analyze the associated MDP. The estimator utilizes a Markov chain Monte Carlo (MCMC) based drone flight path generator and simulator framework which automatically aids in policy adjustments based on failures and improving navigation outcomes.
Funding Agency: DARPA
In Press: [10]
GitHub Repo: github.com/SubMishMar/SNARE_PAPER​
Hybrid CNN-LSTM based Indoor Pedestrian Localization with CSI Fingerprint Maps
Role: Senior Researcher
We present a novel indoor localization system that achieves unprecedented accuracy using Wi-Fi Channel State Information (CSI) fingerprinting. Our key innovation, the 'CSI Fingerprint Map', combines with a hybrid CNN-LSTM architecture to effectively learn spatial-temporal patterns in Wi-Fi signals. Using only standard Wi-Fi infrastructure, our system achieves remarkable accuracy with average errors of just 17cm in static and 36cm in dynamic environments, significantly outperforming existing methods. Through adaptive sliding windows and particle filter-based refinement, we demonstrate that reliable, fine-grained Wi-Fi localization is achievable with minimal infrastructure requirements and moderate environmental noise.
Funding Agency: NSF
In Press: [19]
GitHub repo: github.com/em22ad/CSI-Fingerprint-Indoor-Localization
Monocular SLAM for LNG Plant Inspection
Role: Senior Researcher
For an autonomous mobile robot navigating in an LNG plant, pose estimation is a significant challenge. In this project, we aimed to develop a low cost pose estimation and mapping system by advancing the state-of-the-art in visual localization and mapping, or visual SLAM. A robust and scalable approach enabled a vehicle to operate for long periods of time in a plant environment without requiring reliable GPS reception. We adapted techniques for pose estimation and loop closure were tailored to the unique scenery found in industrialized environments such as LNG plants. The project also captured a range of datasets with ground truth throughout the field tests on a robot platform.
Funding Agency: NSF
GitHub repo: https://github.com/em22ad/
Pan-Tilt Stereovision-based Navigation System for Mobile Robots
Role: Primary Researcher
I designed and implemented an intelligent motion system for mobile robots that uses pan-tilt enabled stereovision system for sensing the environment. The proposed system uses stereovision as a primary method for sensing the environment, and the system is able to navigate intelligently in an indoor environment with varying degrees of obstacle complexity. It creates noiseless and high-confidence 3D point clouds and uses these point clouds as an input for the mapping and path-planning modules.
Funding Agency: National Plan for Science and Technology (KSA)
Vision based Soccer ball FIFA standards Conformation
Role: Project Investigator
An industry-sponsored prototype development was carried out for a vision-based soccer ball FIFA standards conformation system. The project involved detection and localization of defective areas on the soccer ball surface, image segmentation and Hough Transform of ROI for retrieving soccer ball shape from multiple RGB+D cameras, adaptation of Watershed Algorithm and usage of image noise reduction & DIP techniques.
Funding Agency: National University of Sciences and Technology, Pak. AKI Sialkot.
Estimation of Pedestrian Distribution and Tracking via Particle Filter
Role: Primary Researcher
Objective: I developed a highly accurate system for estimating pedestrian distribution in indoor spaces, particularly in crowded and occluded environments. Key Features included, multiple pedestrian detection using laser measurements, particle filter-based tracking algorithm, probability density function estimation for pedestrian counting
I addressed challenges of limited sensors, heavy occlusion, and real-time execution
I implemented a novel approach to measure and account for tracking errors and achieved at least 90% average pedestrian counting accuracy. Algorithm was tested at an exhibit at California Science Center with average pedestrian density of 0.34 per square meter. A C implementation of robust system capable of estimating pedestrian distribution in complex indoor environments was created. It has potential applications in crowd management, space optimization, and security. The findings were presented at ICRA 2009.​
Funding Agency: NSF
Publication: [18]
GitHub Repo: github.com/em22ad/High_Density_People_Detection_Tracking
