Brian Daniel Thomas

AI-ML Intern

UST Blueconch•  February 2023 - April 2023

Worked on Person Re-Identification where a deep learning-based approach to reidentify individuals in CCTV footage was used. The approach utilizes features other than faces and addresses challenges like multiple cameras, angles, occlusions, and lighting. Privacy concerns are addressed by blurring faces in the dataset. The ResNet50 CNN Model achieves an 85% accuracy with high true positive and true negative rates. Given an input image, the model identifies the top 10 most similar individuals for reidentification. Also started some research and work on the usage of SANet (Self-Attention Network), which uses 2 encoders, one is for reducing the distance of errors between the predictions and actual corresponding correct images and the other encoder is for increasing the error distance between the input image and all the other non-similar images. Also did data cleaning, manipulation and created Dashboards on PowerBI