Thomas Daniel Ardhali

Indian Institute of Technology, Kharagpur

M.Tech · Vision and Intelligent Systems

August 1, 2024 – June 30, 2026

Maharaj Vijayaram Gajapathi Raj College of Engineering (Autonomous)

B.Tech · Electronics and Communication Engineering

August 1, 2020 – June 30, 2024

Sri Chaitanya Junior College

Intermediate Education · Andhra Pradesh

June 1, 2018 – May 31, 2020

Jawahar Navodaya Vidyalaya

Secondary Education (Xth)

June 1, 2018 – May 31, 2019

AI-Powered Extreme Weather Vision, Obstacle Detection & Pilot Assistance System

January 1, 2025 – June 1, 2026

Developed a real-time multimodal computer vision system for railway and mining environments by integrating RGB cameras, thermal cameras, and radar sensors for robust obstacle detection under fog, dust, rain, glare, and low-visibility conditions. Built a YOLO-based object detection pipeline in PyTorch and OpenCV for pedestrian and obstacle detection under motion blur and long-range operational scenarios. Implemented semantic segmentation and virtual-fence logic for railway-track understanding, intrusion monitoring, and automated alert generation. Designed OpenCV-based camera calibration, homography estimation, and multi-sensor synchronization pipelines to improve localization accuracy and system reliability. Optimized preprocessing, inference, and sensor fusion pipelines for low-latency real-time deployment and field robustness on embedded edge hardware. Deployed optimized inference pipelines on NVIDIA Jetson AGX Orin 64GB using TensorRT acceleration, FP16 optimization, and asynchronous video processing workflows. Profiled runtime latency, GPU utilization, throughput, and memory efficiency across multiple deployment configurations for real-time edge inference.

Smart Retail Loss Prevention and Queue Analytics System

January 1, 2025 – June 1, 2026

Developed a real-time retail video analytics system for customer tracking, queue monitoring, and suspicious activity detection using multi-camera surveillance streams. Built low-latency object detection and multi-object tracking pipelines using YOLO-based detection models and ByteTrack for persistent identity tracking across crowded scenes. Implemented queue-length estimation, wait-time analytics, intrusion detection, and abandoned-object monitoring using temporal motion analysis and region-based event logic. Optimized inference pipelines using ONNX Runtime and TensorRT with Dockerized FastAPI deployment for scalable GPU-backed real-time inference. Integrated RTSP video streaming, asynchronous frame processing, inference logging, and experiment tracking using MLflow and Weights & Biases for production-oriented deployment monitoring.

M.Tech Thesis Deep Image Restoration for Real-World Lens Flare Suppression

August 1, 2024 – March 1, 2026

Designed MFR-Net (Multi-scale Frequency-aware Residual Network) for real-world lens flare removal using FFT-based frequency attention, half-instance normalization, multi-scale residual learning, and physics-guided artifact subtraction. Benchmarked performance against Restormer, Uformer, HiNet, MPRNet, and U-Net on Flare7K++ and FlareReal600. Engineered UltraLightIRGhostUNet specifically for edge deployment using Ghost Convolutions, depthwise-separable decoding, IRTiny blocks, and ECA attention. Applied INT8/FP16 quantization, mixed precision training (AMP), and knowledge distillation to reduce model size and latency while preserving reconstruction quality. Evaluated using PSNR, SSIM, LPIPS, BRISQUE, and NIQE.

OCR and Handwritten Digit Recognition Pipeline

August 1, 2024 – October 1, 2025

Designed a ResNet9-based deep neural network for handwritten Optical character recognition. Curated a custom Telugu numeral dataset of 1,000+ images with diverse handwriting styles, demonstrating end-to-end dataset preparation and curation. Applied transfer learning using MNIST pretraining and fine-tuning on the custom dataset, achieving strong training and validation accuracy.

Real-Time Object Detection and Classification Pipeline

January 1, 2024 – January 1, 2025

Built an end-to-end object detection and image classification pipeline using YOLO and CNN-based architectures in PyTorch. Prepared annotated datasets using LabelImg and COCO-format workflows with class-balanced augmentation including flipping, color jitter, and mosaic augmentation. Performed preprocessing, cleaning, analysis, training, and validation using OpenCV, NumPy, and Pandas. Evaluated performance using mAP, IoU, precision, recall, and failure-case analysis under occlusion and low-contrast conditions.

Optical Flow Estimation and Motion Analysis in Video

January 1, 2024 – January 1, 2025

Estimated dense motion fields using block matching and Lucas-Kanade methods; fine-tuned FlowNetC and FlowNetS on FlyingChairs. Used OpenCV and PyTorch for frame processing, motion estimation, and flow-map generation. Built a custom video dataset with RAFT-generated pseudo ground truth for real-world motion analysis.

JPEG Compression Codec From-Scratch C++ Implementation

January 1, 2024 – January 1, 2025

Implemented a fully functional JPEG encoder/decoder in C++ covering RGB to YCbCr conversion, 8×8 block DCT, quantization, zigzag scan, and Huffman entropy coding. Validated compression fidelity using PSNR and SSIM across multiple quality factors.