An Empirical Study for Human Behavior Analysis

An Empirical Study for Human Behavior Analysis

Jia Lu (Auckland University of Technology, New Zealand), Jun Shen (Auckland University of Technology, New Zealand), Wei Qi Yan (Auckland University of Technology, New Zealand) and Boris Bačić (Auckland University of Technology, New Zealand)
Copyright: © 2018 |Pages: 18
DOI: 10.4018/978-1-5225-5204-8.ch095
OnDemand PDF Download:
$30.00
List Price: $37.50

Abstract

This paper presents an empirical study for human behavior analysis based on three distinct feature extraction techniques: Histograms of Oriented Gradients (HOG), Local Binary Pattern (LBP) and Scale Invariant Local Ternary Pattern (SILTP). The utilised public videos representing spatio-temporal problem area of investigation include INRIA person detection and Weizmann pedestrian activity datasets. For INRIA dataset, both LBP and HOG were able to eliminate redundant video data and show human-intelligible feature visualisation of extracted features required for classification tasks. However, for Weizmann dataset only HOG feature extraction was found to work well with classifying five selected activities/exercises (walking, running, skipping, jumping and jacking).
Chapter Preview
Top

In surveillance, we describe an event in six facets, namely, What, When, Who, Why, Where and How (5W1H) that could be generalized to feature any surveillance events (Westermann & Jain, 2007). Aligned with studies in video mining and video retrieval (Dai, Zhang, & Li, 2006) (Geetha & Narayanan, 2010), events are regarded to consist of these six 5W1H major components in event recognition and modelling (Xie, Sundaram, & Campbell, 2008). In computing, visual event represents an action or occurrence that could be quantified and recognised by (computing) machine. Similarly, the definition of an event in this paper is the occurrence of something at a particular time and at specific location. In order to facilitate a computer to record, index and arrange video events for users’ post-analysis, the events have a number of attributes including ID, time, location and description. According to the attributes, an event is detected and classified into different classes from the videos in surveillance. Based on categories of an event, we can group the detected events as normal and abnormal ones. For example, Figure 1 shows a normal and abnormal event. Normally, a pedestrian should walk in standing position as in Figure 1(a) or when the walker/bystander falls down as in Figure 1(b) the abnormal event should be detected, and a surveillance alarm should be generated correspondingly and automatically.

Complete Chapter List

Search this Book:
Reset