Mining School Surveys for Quality Education

Conducted Multi-Hypothesis test , to find significance of feedback information from students (20 GB data). Found clusters of area codes based on feedbacks via Locality Sensitive Hashing using Spark+HDFS. Train models on feedbacks via Deep Learing using Tensorflow Distributed Training. Cross nation area codes belonged to same clusters with Jaccard Similarity of 0.80.

Covid-19 Vizualization

Developed a dashboard with Map visualization to show COVID-19 statistics for each country using color-maps. Used parallel coordinate and radar chart to show trends between health expenses, population density and covid-19 stats. Technology used - d3.js, Flask, Bootstrap, jQuery

Sharded Replicated KeyValue Store

Implemented a key value store replicated across multiple machines with RAFT consensus for fault tolerance. The system also used sharding and snapshots for performance. Client sent their requests to KV Server. KV Servers requested raft server for replication, when applied, they replied to client for consistency. This project was developed in GO.

Copter QL: The Q-Learning Helicopter Game

Aimed to make agent learn to play copter using deep reinforcement learning techniques. Implemented a Deep QNetwork (DQN) for learning Q-values for approximate state-action pairs. Exploiting the visited state-action pairs with proper exploration new state-action pairs. Exploring safety conditions to ensure reasonable system performances.

Deep Learning based model to generate nuclei masks from Microscopic Images

Aimed to prepare nuclei masks of microscopic images of cells in bunch. A deep convolutional neural network is modeled, where the features of previous layers are used in future layers, developing a U-Net. Images are first preprocessed with erosion-dilation, then fed into the network, and the output is again processed with watershed labelling, to produce required results, with accuracy metric to be around 47%.

Centrality Metrics for Dynamic Networks

Advisor: Dr. Joydeep Chandra

Formulated centrality metric for dynamic networks, where nodes join or leave the network over course of time. A new hybrid centrality metric is proposed, consisting of PageRank, average importance over time with aging factor also. Citations network is used as the dataset. Metrics obtained for important publications were comparably higher

Adaptive Object Tracking

Advisor: Dr. Jimson Mathew

Used descriptor-based object information [HOG] and condensation algorithm to track people through the frames of video. It was tested on PET 2009 dataset for evaluation, with accuracy around 93%. It is also capable of tracking a person through different cameras in surveillance system. It can also switch the video feed over to another camera whenever a person’s face aligns away towards other camera, providing it with the multi viewing angle.

Lecture Assistant: Federated Human Tracking for Multi-angle Videography

Advisor: Dr. Jimson Mathew & Dr Arijit Mondal

Aimed to assist the lecturer and students in classrooms. An IoT bot was designed having multi angle broad view to track the lecturer and record the lecture in the classrooms using multiple Raspberry Pi. The multi-angle camera is operated by the Raspberry Pi and video is streamed over the server on other Raspberry Pi where students can discuss doubts without interrupting the class. The server also had a quizzing module. This project was demonstrated at ISED Grand Challenge ‘16 and stood runner up amongst others.

Road Traffic Congestion Sensing

Advisor: Dr. Rajiv Misra

Aimed to ease the measurement of road traffic using cheap technology and feasible implementation. Framework has a client end app/senor to send data to server at trigger locations on road. At the server end, after data processing, proposed Map-Reduce engine is used to obtain required road traffic measurements such as average speed, peak hours for each segment of road. This work was accepted for publication at Information Processing in Sensor Networks (IPSN '15).