Browsing by Author "Fadhlan Hafizhelmi bin Kamaru Zaman"

Recent research in video surveillance system has shown an increasing focus on creating reliable systems utilizing non-computationally expensive technique for observing humans' appearance, movements and activities, thus providing analytical information for advanced human behavior analysis and realistic human modeling. In order for the system to function, it requires robust method for detecting and tracking human from a given input of video streams. In this thesis, a human detection technique suitable for video surveillance is presented which requires fast computations in addition of accurate results. The techniques proposed include adaptive frame differencing for background subtraction, contrast adjustment for shadow removal, and shape based approach for human detection. The tracking technique on the other hand uses correspondence approach. Event Based Video Retrieval (EBVR) system is also proposed for efficient surveillance data management and automated human recognition with unique ID assignment. Proposed human detection and tracking are integrated with EBVR and motion detection into a complete automated surveillance system called Active Vis Video Surveillance Analysis System (AVSAS) which produces good result and real-time performance especially in non-crowded scene. The EBVR system also proves to be able to handle automated human recognition with unique ID assignment accurately.

Conventional face recognition methods usually assume the possession of multiple samples per person (MSPP) available for classification. This assumption however, may not hold in many practical face recognition applications since only single sample per person (SSPP) is available for enrollment. The scarcity in numbers of training sample could deteriorate the reliability of many popular face recognition methods. Thus, in this thesis, a novel semi-supervised face recognition approach is proposed to address the SSPP problem by effectively extracting the inherent information in face image through local ensembles strategy. A Spiking Neural Network (SNN) based classifier called Coincidence Detection SNN (CD SNN) is proposed which identifies the synchronization between input spikes and at the same time employs the psychophysically-relevant feature selection through synaptic time constant prediction (?_(s )Prediction) as bias for more accurate face classification. The CD SNN classifier is built on top an improved Zero-Order Spike Response Model (SRM0), utilizing spike time approximation using the proposed Output Spike Time Prediction (OSTP) approach for faster computation. The classification is then performed on more efficient and compact image representations acquired through SNN Face Descriptor (SNN FD). Comparisons with several state-of-the arts methods using several popular face datasets reveal that the proposed method can achieve equivalent performance under SSPP constraints, and in fact on several occasions, delivers significantly better performance than existing methods. Additionally, through a survey, it is found that proposed method performs better than human in SSPP face recognition. Based on the same survey, assessment on the difference of feature selection between human and proposed method is also presented.