Date of Award

31-8-2024

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.D.Manivannan

Keywords

Artificial Neural Network, Context-Free Grammar, Rare Event Detection, Hypergraph

Abstract

Abnormal event detection aims to identify the events that deviate from expected normal patterns. This work primarily focuses on detection of rare events in public places. The existing research challenges in a video-based surveillance systems for the vehicle have been analysed and presence of abnormal objects in traffic-oriented videos have been detected. A novel approach for event summarization and rare event detection has been proposed in this work. The key ingredient in this work is the incorporation of Hypergraph (HG) matching.

Despite the reasonable amount of success achieved by a large number of researchers over the globe, distinguishing important videos from recorded information and labelling the important videos in an appropriate order for detecting a particular rare event continues to be a challenging task. The challenges are further compounded in real life applications where relevant, nonredundant, and timely event summaries are needed for applications. Unlike conventional methods that focus on a specific concept and approach, the proposed work deals with a Hybrid Model.

Initially, a Context-Free Grammar (CFG) is developed for detecting rare events from a video stream and Artificial Neural Network (ANN) is used to train CFG. A set of dedicated algorithms are used to perform frame split process, edge detection, background subtraction and convert the processed data into CFG. As an outcome, we have a single object detection scheme from labelled events in the given video frames, limitation being not amenable for generalization.

A group of preprocessing feature matching algorithms were experimented and it was found that modified KAZE provided the highest accuracy. As feature extraction is a key requirement, all aspects of the object must be considered. Hypergraph provides an exciting computational facility to extract both geometrical and topological features simultaneously and hence helpful to quickly identify rare events at lower computational complexity. Hypergraphs can model higher order data.

Hypergraph clustering refers to the process of extracting maximally coherent groups from a set of objects using higher order similarities. This work deals with one novel Hypergraph Matching Algorithm integrated with modified KAZE key-points and Feature extraction technique to process video frames for tracking multiple objects.

During processing of images, it could be observed that analysing each pixel in innumerable number of videos is practically impossible for a human being and highly computationally inefficient thereby advocating the need for sampling. Grid sampling using Helly property has been found to be the most effective increasing the accuracy of modified KAZE further. The novel feature in this approach deals with its ability in handling outliers.

Further improvements are made in order to handle ambiguities with multiple feature descriptors. To incorporate the same, a hypergraph clustering algorithm has been viewed in terms of a noncooperative multiplayer clustering game. During each iteration, the image needs to be reconstructed. For accurate reconstruction, a convolutional autoencoder has been used. This autoencoder is found to be the most suitable while handling video data. The novelty in this encoder is its ability to capture low-level correlation between spatial and temporal dimensions of videos and also generating distinctive features representing visual spatio-temporal information.

Experimental results and comparison with the state-of-the-art methods have demonstrated the effectiveness of our proposed method in detecting abnormal events in a faster manner than the recently designed algorithms.

Overall HG based rare event detection in videoframes requires matching algorithms, Game theoretical concepts and Machine learning algorithms such as Autoencoder and Convolutional Neural Network for quick processing of videos.

Download

Share

COinS