Khulna University of Engineering & Technology

Multidimensional arrays are greatly used for handling large amount of data in scientific or engineering, and Database applications. Most of the on hand data structures are static in nature. We describe a novel implementation idea of multidimensional array for handling such large scale datasets. The scheme implements a dynamic multidimensional extendible array employing a set of two dimensional extendible arrays. The Traditional Multidimensional Array (TMA) or Extended Karnaugh Map Represented (EKMR) array is an efficient structure in terms of accessing the element of the array by straight computation of the addressing function, but they are not extendible during run time. But real world data grows in incremental fashion. So, there is strong demand of data structure that is dynamically extendible during run time. Three are some extendible array models, most of which uses a concept of extension subarray. For n-dimensional array the subarrays are n-1 dimensional. But, if the length of dimension and/or number of dimension of a multidimensional array is large then the address space, even for the subarray, overflows the machine limit very soon. Another issue for representing the real life data by multidimensional arrays is that it creates a problem of high degree of sparsity and need to be compressed. It is therefore desirable to develop techniques that can access the data in their compressed form and can perform logical operations directly on the compressed data. In this research work we propose a data structure using the idea of EKMR and Traditional Extendible Array, namely Extendible Karnaugh Array (EKA) to represent the multidimensional data. The scheme has the intuitive propensity against the essential problem of address space overflow as well as it can be extended in any direction during run time. Moreover, we present a compression scheme for EKA to facilitate data access in compressed form. We evaluate our proposed scheme by comparing for different retrieval and extension operations with the Traditional Multidimensional Array (TMA). Our experimental result shows that the EKA scheme has a significant delay on the occurrence of address space overflow without any performance penalty. Furthermore, we find that range of usability of the compression scheme is independent of length or number of dimension. And it is better to use compressed EKA rather than uncompressed EKA for representing sparse data sets which needs range retrieval frequently.

In recent years, analyzing human motion and recognizing a performed action from a video sequence has become very important and has been a well-researched topic in the field of computer vision. The reason behind such attention is its diverse applications in different domains like robotics, human computer interaction, video surveillance, controller-free gaming, video indexing, mixed or virtual reality, intelligent environments, etc. There are a number of researches performed on motion recognition in the last few decades. The state of the art action recognition schemes generally use a holistic or a body part based approach to represent actions. Most of the methods provide reasonable recognition results, but they are sometimes not suitable for online or real time systems because of their complexity in action representation. In this thesis, we address this issue by proposing a novel action representation scheme.The proposed action descriptor is based on a basic idea that rather than detecting the exact body parts or analyzing each action sequence, human action can be represented by a distribution of local texture patterns extracted from spatiotemporal templates. In this study, we use a novel way of generating those templates. Motion History Image (MHI) merges an action sequence into a single template. However, having the problem in overwriting old information by a new one in the MHI, we use a variant named Directional MHI (DMHI) to diffuse the action sequence into four directional templates. And then we use the Local Binary Pattern (LBP) operator, but with a unique way, a rotated bit arranged LBP, to extract the local texture patterns from those DMHI templates. These spatiotemporal patterns form the basis of our action descriptor which is formulated into a concatenated block histogram to serve as a feature vector for action recognition. However, the extracted patterns by LBP tends to lose the temporal information in a DMHI, therefore we take a linear combination of the motion history information and texture information to represent an action sequence. We also use some variants of the proposed action representation that include the shape or pose information of the action silhouettes as a form of histogram.We show that, by effective classification of such histograms, i.e., action descriptor, robust human action recognition is possible. We demonstrate the effectiveness of the proposed method along with some variants of the method over two benchmark dataset; the Weizmann dataset and KTH dataset. Our results are directly comparable or superior to the results reported over these datasets. Higher recognition rates found in the experiment suggest that, compared to complex representation, the proposed simple and compact representation can achieve robust recognition of human activity for practical use. Besides the recognition rate, due to the simplicity of the proposed technique, it is also advantageous with respect to computational load.