Building intelligent vision and data-driven systems.

Introducing My Research Journey at the UD Vision Lab: Advancing Object Detection and Computer Vision.

Created the DAVIS dataset to support vehicle detection and behavior analysis research at the UD Vision Lab. This dataset provides training data for YOLO-based object detection systems in roadside restricted-area monitoring.

• Curated and annotated USA road traffic footage with cars, trucks, buses, and pedestrians.
• Generated high-quality YOLO-format bounding box labels for all object instances.
• Captured real-world scenarios including occlusion, night scenes, motion blur, and variable angles.
• Used extensively in our driving-behavior analysis system and trajectory prediction experiments.

Implemented a full MATLAB simulation of a 4f optical system that synthesizes a complex Fresnel hologram using a single amplitude-only SLM. Decomposed the hologram into real and imaginary components, applied Fourier-plane grating modulation, overlapped ±1 diffraction orders, and reconstructed the object using Fresnel back-propagation. The method replicates the technique proposed by Liu et al. (Applied Optics, 2011).

Designed a modular pipeline using SAM segmentation + DINOv2 feature encoding and a 3-layer classifier to detect unattended tools and equipment across 500+ real construction site images, achieving 92% bounding box precision and 87% validation accuracy across 20 hazard classes.

Developed a hierarchical segmentation pipeline to differentiate foreground hazardous objects from background surfaces in construction scenes. Utilized a two-stage Segment Anything Model (SAM) approach with morphological processing, connected-component merging, and area-based filtering to accurately isolate objects such as tools, wires, and construction equipment. Implemented a 15% region-area threshold to classify potential hazards and demonstrated reliable detection of falling-object and caught-in/between risk scenarios.

Developed an automated traffic monitoring system to analyze driving behavior in roadside restricted zones, processing 30 fps streams through YOLOv8. Improved low-light detection using R-CLAHE and forecasted trajectories 3–5 seconds ahead with Kalman filters, reducing false alarms by 60%.

Implemented a classical pedestrian detection pipeline using Histogram of Oriented Gradients (HOG) for feature extraction and a binary Support Vector Machine classifier. Processed 128×64 image patches, generated 3,780-dimension feature vectors per window, and trained an SVM on positive/negative pedestrian datasets, achieving 97.71% accuracy with low false-positive and false-negative rates.

Implemented a Hopfield-based associative memory system to restore corrupted MNIST digits. Normalized and binarized images to ±1 states, constructed a recurrent weight matrix using Hebbian learning, and iteratively updated neuron states using the signum activation rule. Successfully reconstructed noisy inputs (5–25% salt-and-pepper noise) and achieved stable recall for orthogonal pattern sets up to 82% noise levels.

Built an SRGAN model to enhance satellite image resolution by 4×, improving PSNR by 3.2 dB and SSIM by 0.15 over bicubic interpolation. Designed the full training pipeline from data preprocessing to evaluation.

Deployed an IoT monitoring system using Arduino and ESP32 to collect sensor data every 30 seconds from 4 environmental parameters, triggering automated irrigation and climate control and reducing manual monitoring by 90%.

Developed an animal image classification system using handcrafted feature extraction and neural classification to distinguish between 32 animal species, targeting applications in wildlife monitoring and livestock management.

• Used a dataset of 32 animal classes (e.g., antelope, bear, elephant, tiger, wolf, zebra) with 50 training and 10 testing images per class (1600 train / 320 test images).
• Extracted Local Binary Pattern (LBP) features from resized 64×128 grayscale images, followed by Histogram of Oriented Gradients (HOG) for richer texture–shape descriptors.
• Trained a Multilayer Perceptron (MLP) with 3780 input features, 50 hidden units, and 32 output neurons to classify species based on the combined feature representation.
• Evaluated performance using true positive and false positive rates per class, visualized via comparative bar graphs to analyze strengths and confusion across species.

Launched an Android app with 500+ downloads for real-time bus tracking and schedule visualization, reducing average commuter wait time by 12 minutes based on a 200-user survey.

Designed and implemented frequency-domain lowpass filters using FFT2/IFFT2 to analyze and suppress high-frequency noise in coherent imaging systems. Built both Ideal and Gaussian lowpass filter masks based on Euclidean distance in the frequency plane, with cutoff frequency Fc = 50 and Gaussian σ = 50. Demonstrated the effects of sharp vs. smooth transitions in frequency filtering, including ringing artifacts in ideal filters and improved coherence preservation in Gaussian filtering.

Designed a scalable big-data analytics pipeline using Apache Hive, Spark, and MapReduce to process 100,000+ Indian aviation records on a Cloudera Hadoop cluster, enabling insights into national flight traffic patterns and operational performance.

• Built end-to-end ETL pipeline using HiveQL joins across 5+ dimensional tables (flights, airports, carriers, delays, passengers) to extract actionable aviation intelligence.
• Implemented Spark SQL aggregations and MapReduce jobs to analyze the top 15 busiest domestic destinations, passenger load factor distribution, and delay patterns by route, carrier, and time-of-day.
• Identified that 23% of flights experienced more than 30-minute delays during Indian peak travel hours.
• Optimized Hive performance using table partitioning and MapReduce job tuning, reducing average query runtime from 45s to 12s (73% improvement) on multi-terabyte datasets within the Hadoop cluster.