Lists
Lists
Afzal, Muhammad Majid (2019) Classifier based incremental reconstruction of human object trajectory in live video streams. Doctoral thesis, London Metropolitan University.
A common practice for solving problems in video analytics is to consider each problem (e.g. processing speed, complexity of data analysis or tracking of moving object) in isolation, starting with constructing of computational models for each specific data processing task. This is dictated by the need to address the complexity of the overall problem, which typically can be broken into multiple computational tasks - detection, recognition, classification, analysis, prediction, etc. Each of these tasks has its own computational model and uses different data. This approach has been the main focus of research in video processing for a long time and had some success, but it leads to the significant requirement for computational power and is not feasible for real-time video analytics tasks because of the time constraints.
An alternative to this strategy is to link the video analytics tasks in order to leverage the related data sets and computational models. By doing that we can decrease the amount of data and computations necessary and thus, reach satisfactory performance. Consequently, we can improve the accuracy of the solution obtained by simplification through utilizing the previous experience through learning. Our aim in this research is to exploit the variety of parameters of this process - development of the data processing model, features extracted from the data and algorithms of the machine learning.
This research addresses the problem of reconstructing moving object trajectories for real-time video analytics based on continuous processing of input video streams with application to a wide range of tasks in video surveillance. Our approach uses a combination of classifiers for different features of the data, such as shape and colour, as well as other parameters (e.g. the position of human body parts) which can be used for detecting various profiles of interest for further processing of the original video stream. The custom-tailored model of moving object is created by experimenting with different models and algorithms specifically applicable to the task for trajectory reconstruction and selected to make the best use of data and to achieve the best result in terms of speed of data processing which is further discussed in chapter 6.
On the basis of the experimental work conducted in this research we have developed a software system, which can track and identify moving human objects of interest within live video streams through a combination of visual pre-processing (discussed in chapter 5 in video transformation module), machine learning and video post-processing (discussed in chapter 5 in ‘rotation’ activity) methods. It uses a novel object detection algorithm which significantly improves the speed of processing during the tracking and localization steps. This algorithm makes use of a number of comparing analytics method by classifying the shapes and sizes parameters as shown in chapter 4 with different cases.
Although specifically designed to serve the purpose of the research programme for behaviour analysis of moving objects in live video streams this part of our framework can be considered as a standalone problem solving module for software systems which involve object detection, identification and tracking based on live video input. It can be used as an open source software in many areas ranging from public safety management and video surveillance to computer games and animation.
Available under License Creative Commons Attribution Non-commercial No Derivatives 4.0.
Download (5MB) | Preview
 |
View Item |