This paper describes an automatic lip reading system consisting of two main modules 1) a pre-processing module able to extract lip geometry information from the video sequence and 2) a classification module to identify the visual speech based on dynamic lip movements. The recognition performance of the proposed system has been assessed in the recognition of the English digits 0 to 9 as spoken by the speakers in the video sequences available in the CUAVE database. Extraction of lip geometry features was carried out using a combination of a skin color filter, a border following algorithm and a convex hull approach. The proposed method was compared with the popular `snake' technique and was found to improve lip shape extraction performance for the database studied. Lip geometry features including height, width, ratio, area, perimeter and various combinations of these features were evaluated to determine which performs the best when representing speech in the visual domain in the application of three separate classification methods, namely optical flow, Dynamic Time Warping (DTW) and a new approach termed Multi-Dimensional DTW. Experiments show that the proposed system is capable of a recognition performance of 68% just using lip height, lip width and the ratio of these features demonstrating that the system has the potential to be incorporated in a multimodal speech recognition system for use in noisy environments.

This paper introduces a genetic algorithm (GA) planner that is able to rapidly determine optimal or near-optimal solutions for mobile robot path planning problems in environments containing moving obstacles. The method restricts the search space to the vertices of the obstacles, obviating the need to search the entire environment as in earlier GA-based approaches. The new approach is able to produce an off-line plan through an environment containing dynamic obstacles, and can also re-calculate the plan on-line to deal with any motion changes encountered. A particularly novel aspect of the work is the incorporation of the selection of robot speed into the GA genes. The results from a number of realistic environments demonstrate that planning changes in robot speed significantly improves the efficiency of movement through the static and moving obstacles.

This paper introduces a novel genetic algorithm whose properties have been purposely designed to be suited to hardware implementation. This is distinct from previous hardware designs that have been realized directly from conventional genetic algorithm approaches. To be suitable for hardware implementation, we propose that a genetic algorithm should attempt to both minimize final layout dimensions and reduce execution time while not compromising algorithmic performance. Consequently, the new genetic algorithm specifically aims to keep the requisite silicon area to a minimum by incorporating a monogenetic strategy that retains only the optimum individual, resulting in a dramatic reduction in the memory requirement and obviating the need for crossover circuitry. The results given in this paper demonstrate that new approach improves on a number of existing hardware genetic algorithm implementations in terms of the quality of the solution produced, the calculation time and the hardware component requirements.

Incremental learning is vital in robots that need to adapt rapidly and respond appropriately to new or unexpected events that occur in its vicinity. This paper describes a novel incremental learning method here demonstrated in its reactive navigation of a mobile robot. The approach records acquired feature vectors in a compact statistical representation we term a frequency table that is purposely designed to contain sufficient information for the generation of a conventional C4.5 decision tree. During learning, further feature vectors can be added in an incremental manner to the frequency table which can then be used to generate new decision trees as required. A major advantage compared with previous incremental decision tree approaches that keep feature vectors in the tree itself, is that the frequency table's memory requirement is fixed and known a priori. Experimental results in the practical application of the method to the real-time navigation of a mobile robot through previously uncharted environments show that the calculation time is typically an order of magnitude less than that of an incremental generation of an ITI decision tree

Logic density increases have made feasible the implementation of multiprocessor systems able to meet the intensive data processing demands of highly concurrent systems. We describe the research and hardware implementation of a high-performance parallel multicompressor chip. A detailed investigation into the performances of alternative input and output routing strategies for realistic data sets demonstrate that the design of parallel compression devices involves important trade offs that affect compression performance, latency, and throughput. The most promising approach is implemented into FPGA hardware and is shown to provide a scalable compression solution at throughputs able to cope with the demands of modern high-bandwidth applications.

Proposes a hybrid technique for intelligent robot navigation based on incremental decision trees (ITI-2.8) and incorporating fuzzy logic for flexible control. The robot perception is decomposed into a hierarchy of simpler virtual environments, termed worlds. Training examples generated from the robot's past rewarded experiences are exposed to ITI-2.8 in an incremental manner and online to evolve an array of fuzzy associative memories (FAM), each representing a unique world. That is, generated FAMs, which are structurally nonlinear (in contrast to ordinary FAMs), are engineered online and from inception to store and access fuzzy control rule spaces representing different perceptions. Each decision tree is encoded in one FAM and is local to a certain perception. The fundamental strengths of the algorithm in building online FAMs, is its incremental nature and automatically generating fuzzy training vectors without human intervention. Fuzziness is integrated to provide suitable reasoning in the face of inherent uncertainty in the sensory input data and to merge conflicting behaviours to generate smooth trajectories. Global navigation is achieved by activating a hierarchy of local FAMs.