Date of Award

3-27-2019

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Systems Engineering

First Advisor

Mariofanna Milanova

Abstract

The human action recognition in video surveillance (HARVS) system is the system that classifies and detects action in videos. The system depends on two modes: training mode, in which the system learns about the human actions in video samples, and testing mode, in which the system recognizes the unknown action in testing video samples. HARVS in computer vision is a desirable approach because of its various desired applications such as surveillance environments, security systems, robot vision machine, human-computer interaction, human traffic accident detectors, multimedia retrieval, entertainment environments, and healthcare systems. The main aim of the work in this dissertation is to build and establish an accurate, efficient, and fast system using different types of techniques. This research intends to study and investigate a system that involves three main stages: 3D video frames extraction, feature extraction, and action recognition processes (classification and detection). To accomplish the main aim, various objectives are categorized in terms of both accuracy and recognition speed. The main contribution of this work is to introduce a new system of human action recognition in video surveillance. A scale, rotation, and translations descriptor based on Hu Moment Invariants (HMI) in feature extraction and Euclidean distance classifier (EDC) for the classification process has achieved the most promising performance. In addition, a new real-time deep neural network model has been used by applying the Caffe_GoogLeNet framework with different training epoch values (TEs) with a remarkably short running time. Next, human health-related video actions are detected using an Android camera based on the TensorFlow Object Detection Application Program Interface (API) technique, with two new per-trained models: SSD-Mobilenet and Faster R-CNN-Resnet. One benefit of using a pre-trained model is that instead of generating the model from scratch, a model that has been trained for a related problem can be utilized as a starting point to train the system. This dissertation is implemented under different camera recording conditions utilizing four different datasets: KTH, Weizmann, UCF101, and NTU RGB+D with gray and color resolutions.

Download

Included in

Engineering Commons

Share

COinS