Person of interest rapid identification and tracking (POIRIT) assistance system

Arif, Danish

Person of interest rapid identification and tracking (POIRIT) assistance system

Thesis / Dissertation

2024

Abstract

Pakistan, situated in a highly militarized neighborhood, has been entangled in persistent conflicts. The repercussions of the Soviet-Afghan and US-Afghan wars led to an influx of millions of refugees into Pakistan, resulting in significant financial and security challenges. In response, Pakistan's armed forces and intelligence agencies successfully cleared affected areas of terrorist presence, ushering in a semblance of peace. However, this progress gave rise to new issues, as terrorist and militant factions covertly infiltrated cities, engaging in both acts of terror and street crime, thereby compromising urban safety. To counter this, the government established a specialized street crime patrol unit known as "Dolphins," equipped with modern surveillance technology, including video feedback systems. Additionally, the "Punjab Safe Cities" initiative was launched, involving the installation of cameras across cities to monitor citizen movements. While the existing system boasts 10,000 operational cameras in Lahore, it primarily relies on manual monitoring, lacking advanced features for proactive threat detection. To address this, an automated solution employing Artificial Intelligence (AI) is proposed, specifically designed to track "High-Value Targets" by processing visual data feeds from the urban environment. It begins with image acquisition from diverse network CCTV cameras across different geographical locations. This involves the implementation of network cameras with controlled frame rates, facilitated by Raspberry Pi nodes and augmented with PIR sensors for efficient frame change detection. The parallel operation of nodes and cloud uploads is integral to enhancing processing speed and reducing computational burden. Furthermore, the study deliberates on resizing images from varying CCTV camera resolutions, a critical step to mitigate computational costs without compromising performance. By employing the DORI rule, the frame resolution is adeptly adjusted, resulting in substantial memory and computational savings. The research also introduces a robust approach to cropping and storing essential facial features, emphasizing the significance of bizygomatic measurements and additional regions to account for tilts. In terms of face detection techniques, the study advocates for the utilization of YOLO v3 due to its superior performance in detecting small face faces from CCTV frames. POIRIT showcases an accuracy rate of 90.7% while detecting small-sized faces in challenging environments which makes it superior to previous studies. The research further delves into the implementation of cascaded classification layers tailored to the demographic of Pakistan, efficiently categorizing images based on gender, face type, and facial hair styles. The study meticulously prepares datasets for training deep learning models, encompassing a diverse range of samples from various urban contexts. The research employs transfer learning models like AlexNet, GoogLeNet, ResNet, and VGG-Face, ultimately determining the most effective model for each classification layer. For gender classification, VGG Face exhibits the best performance with an accuracy of 98.41%, which is 1% more than the best performer in 2021. Likewise, for ethnicity classification, VGG Face demonstrates an overall accuracy of 97.2% proposing a unique work of Pakistan region with higher accuracy than models trained for other ethnicities. The final layer of facial hair type classification reveals that ResNet-50 outperforms other models with an overall accuracy of 97.31%. This novel work can classify various facial hair styles, but its accuracy is higher than the facial hair detection systems available. These results demonstrate the effectiveness of the chosen models in their respective classification tasks. The system not only achieves high accuracy rates but also introduces advanced image enhancement capabilities, offering solutions for low-resolution image estimation and sketch-to-color conversions. It incorporates the use of Generative Adversarial Networks (GANs) to introduce advanced image enhancement capabilities. The results demonstrate the superior performance of MS Cycle-GAN, with lower Mean Square Error (MSE) and Root Mean Square Error (RMSE) values. Moreover, the study addresses database creation and accessibility, ensuring a centralized repository for organized storage and easy retrieval of cropped images. The proposed distributed computing system, leveraging Raspberry Pi, CPUs, and GPUs, is pivotal in parallel data processing. Additionally, the research considers environmental factors like fog and smog, incorporating the Retinex Algorithm with Wavelet transform for effective fog removal. The search for persons of interest is streamlined through radial matching using cosine similarity, streamlining the process, and enhancing efficiency. Overall, the research presents a comprehensive framework for an intelligent assistance system, incorporating cutting-edge techniques and addressing critical considerations for effective facial recognition in urban environments. The system aims to assist security agencies in tracking high-value targets in busy urban areas. It offers customizable solutions with the potential for diverse applications and can be enhanced with newer deep-learning models. This research holds promise for advancing biometrics and parallel computing, adaptable for global use. Further refinements could include additional target subclassifications, activity monitoring, and optimized hardware configurations. Implementing this system in developing countries like Pakistan could lead to safer cities, driving economic growth through increased business activities and tourism. These positive outcomes would not only benefit the nation but also strengthen international ties and promote innovation.

Keywords

Electrical Engineering

Reference:

Collections

PhD / Doctoral

Full item page