A new algorithm for symbol recognition is proposed in this paper. It is based on the AutoClass classifier [1], [2], which itself is a version of the evolving fuzzy rule-based classifier eClass [3] in which AnYa[1] type of fuzzy rules and data density are used. In this classifier, symbol recognition task is divided into two stages: feature extraction, and recognition based on feature vector. This approach gives flexibility, allowing us to use various feature sets for one classifier. The feature extraction is performed by means of gist image descriptors[4] augmented by several additional features. In this method, we map the symbol images into the feature space, and then we apply AutoClass classifier in order to recognise them. Unlike many of the state-ofthe-art algorithms, the proposed algorithm is evolving, i.e. it has a capability of incremental learning as well as ability to change its structure during the training phase. The classifier update is performed sample by sample, and we should not memorize the training set to provide recognition or further update. It gives a possibility to adapt the classifier to the broadening and changing data sets, which is especially useful for large scale systems improvement during exploitation. More, the classifier is computationally cheap, and it has shown stable recognition time during the increase of training data set size that is extremely important for online applications.

No abstract

No abstract

In this article a Bayesian filter approximation is proposed for simultaneous multiple target detection and tracking and then applied for object detection on video from moving camera. The inference uses the evidence lower bound optimisation for Gaussian mixtures. The proposed filter is capable of real time data processing and may be used as a basis for data fusion.The method we propose was tested on the video with dynamic background,where the velocity with respect to the background is used to discriminate the objects. The framework does not depend on the feature space, that means that different feature spaces can be unrestrictedly used while preserving the structure of the filter.