Human Activity Recognition in UAV Videos using Deep Learning Techniques

Abstract

An Unmanned Aerial Vehicle (UAV) is an aircraft that operates without a pilot on newlineboard and possesses the ability to reach inaccessible areas. It captures high-quality im- ages or videos at a reduced cost, revolutionizing various fields such as surveillance, land surveying, media, agriculture, and emergency management. However, the increasing reliance on UAVs for surveillance and investigation calls for precise human detection and tracking. The freedom of flight presents various challenges in this regard. Over the past few years, there has been a significant transformation in vision-based human detection and tracking methodologies, moving away from traditional methods and em- bracing deep learning techniques. Deep learning enables automatic feature learning,and the rich content obtained from UAVs opens up new avenues for exploration. newlineThis thesis introduces effective methodologies for human detection and tracking in UAV videos. A groundbreaking framework for optimizing UAV videos is devel- newlineoped, specifically designed for human tracking purposes. The framework incorporates newlinekeyframe extraction, principal keyframe selection, and human path analysis techniques newlineto achieve highly optimized UAV videos with minimal path deviation. newlineThe initial research presents a neural network framework named Mask-Recurrent newlineNeural Network (Mask-RCNN), which employs advanced deep learning methods to newlineidentify and categorize individuals in aerial imagery data. Moreover, the use of His- newlinetogram Oriented Gradients (HOG) algorithm is demonstrated to localize specific re- newlinegions containing objects within the image. The study also introduces an algorithm newlinecalled histogram-based Mask-RCNN, which forms the core of the proposed framework newlineand governs its overall functionality. newlineThe second study introduces a human action recognition model for UAV videos, newlineutilizing the Inflated I3D-ConvNet (Inflated I3D) and Bidirectional Long Short-Term newlineMemory (Bi-LSTM) approaches to achieve accurate recognition of single human ac- newlinetivities. The initial module of the mod