الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 32 |  |  |
| | | from | 32 | | |
Abstract
Intelligent surveillance aims at conceiving reliable and efficient systems that are able to detect and track moving objects in complicated real world scenes. The behavior of these objects is then analyzed and described. A general framework of visual surveillance systems includes environment modeling, motion segmentation, object classification, tracking, as well as behavior understanding and description. Motion detection is the module responsible of segmenting moving objects from static or irrelevant background. This module is the base for any subsequent processing; thus it must be accurate, robust and fast. Human action classification is also a very important module in visual surveillance that enables for automatic surveillance and analysis of suspicious behaviors.
The contribution of this thesis can be divided into two main components. The first is in the field of motion detection and the second is concerned with human action recognition. First, a method is proposed for enhancing the performance of accumulative frame differencing. This method succeeded in the detection of very small/slow and low contrast objects moving in environments with varying illumination. It also overcomes the aperture problem encountered in frame differencing, but has thedrawbacksof having false positive spots (False Alarms) and that the dynamic background problem is not completely solved. Four proposed variations of this method have been investigated. The first variation gives the best average performance overall studied realistic cases without raising the complexity of the MFD technique.
Also in the field of motion detection, the 3D SWT has been proposed and used for detecting moving objects. The proposedspatio-temporal motion detection techniqueis based on a multi-resolution stationary wavelet analysis. The performance analysis showed that the proposed 3D spatio-temporal technique outperforms classic techniques without increasing the time complexity and gives a compromise between different performance evaluation metrics. The 3D technique proved to be applicable in different applications and scenarios, especially in temporally varying illumination conditions and the applications with small/ slow law contrast objects.
The second main contribution is in the part of human action recognition. Human action and activity recognition are very challenging and interesting fields in computer vision. A large number of challenges face the process of action recognition. Variations in the environment and recording settings andvariations in the performance of the action are examples of such challenges.
Multi-resolution methods and especially wavelets have been extensively used in signal, image and video processing. This is a result of their successful and robust utilization in many applications. The stationary wavelet transform is one of the wavelet transforms that have been employed in some computer vision applications. The SWT gives a better approximation than the DWT since it is linear, redundant and shift invariant.
A new human action descriptor is proposed. The proposed descriptor combines the power of wavelets in highlighting directional variations and the simplicity and robustness of local binary patterns in describing local structures. The new local descriptor is based on the 3D stationary wavelet transform that highlights spatio-temporal variations representing human actions in video sequences. The proposed directional features has been tested on two public datasets and their accuracy has been verified using two classifiers. Results show the power of fusing global and local descriptors in the process of human action recognition. A weighted combination of the directional features has also been investigated.
Future work may include the following directions:
• Using different wavelet functions and investigating the effect of increasing the number of analysis levels, in the temporal analysis part of the proposed separate spatio-temporal technique.
• Using high performance techniques to further speedup the processing of the 3D SWT.
• Investigating the effect of block based processing of the wavelet multi-scale human action templates.
• Extending the new descriptors to the 3D domain.
• Investigating the effect of using different weighting schemes to combine the directional feature vectors.
• Using the extracted global features for tracking the moving objects.