Projects

As a pilot foundational component of the Smart Cities Mission India project, we are developing a digital twin of Bengaluru to enable advanced traffic forecasting and management.

Our approach leverages computer vision to analyze real-time vehicle turning patterns at key intersections, establishing a baseline for traffic behaviour. This data is then enriched using visual AI to re-identify vehicles across a distributed camera network, allowing us to find origin-destination pairs.

By feeding this information into a Geometric Deep Learning model, we can analyze the entire road network's complex dynamics. This robust framework allows us to run predictive "what-if" scenarios—such as modelling the city-wide impact of a sudden road closure. The ultimate goal is to accurately determine the origin-destination (OD) pairs for vehicles, providing actionable intelligence to optimize traffic flow, reduce congestion, and urban planning

Let's get fit! We developed an end-to-end mobile application to track the calorie intake vs calories burnt. It provides a net calorie per day with some other interesting statistics. You can even click an image of a fruit/vegetable (currently supporting 36 fruits and vegetables) and our image classification algorithm gives back the total calories per 100 grams from the item clicked.

GitHub - https://github.com/parikshit14/my-plate-nutrition 

Capture to extract information from a bill! Leveraging the power of OpenCV for image processing we developed an automated software that extracts the information from the clicked image of a bill. The GIF showcases the steps with the evolution from camera capture to skew correction to contour detection and at the end information extraction using Optical Character Recognition. This extracted information like the date, bill total, etc., can be further extracted to Excel sheet to get your monthly expenses sorted.

GitHub: https://github.com/parikshit14/exp-manager 

Rock, Paper, Scissor, Shoot! This is a stone-paper-scissors object detection model using the YOLOv3 architecture. Right from the dataset preparation, to model construction it was done manually by me. Object localization using YOLO is generally preferred because of the speed, accuracy and training efficiency over other models. 

GitHub: https://github.com/parikshit14/ImplementingYOLO  

Empowering Precision Farming: Pioneering Lighter AI Models for Faster, Accessible Agriculture! The Project was focused on making precision farming more reachable. What generally happens with AI models is that they require high computational resources like GPUs, TPUs to function in real time and provide accurate results. To mitigate this issue, a lighter 8-bit integer model is being used because it turns out the precision of 32-bit floating point is not needed when we consider inference. This in turn reduced the processing time by a factor of six and model size by a factor of 4. The models are tested on RPi, Phone and Laptop processors for a comprehensive performance comparison.

GitHub: https://github.com/parikshit14/QNN-for-weed