Feature vectors for road vehicle scene classification

Abstract

Effective feature vectors are needed if artificial neural networks (ANNs) are to provide an efficient method for classifying an area of interest (AOI) in a visual image captured from closed circuit television systems (CCTV) with the camera static. Problems of object orientation (rotation) and scale (size variation due to viewing distances) in an image are important in moving object classification or recognition. In this paper, feature vectors for the subimage moving objects associated with urban road scenes have been derived from invariant geometric properties of the object and statistical parameters of the image scene. These have been used as inputs for a neural network classifier. The extension of binary moment invariant techniques to grey level moment invariants has provided enhanced feature vectors for classification of objects in such images and shown to be effective. Anomalies such as moving shadows, swaying trees, changing reflectivity, etc., could be overcome by suitable preclassification neural stages. It has been shown that incorporating these anomalies along with the objects, resulting in a one-stage system, can give the best result, particularly when the training data for anomalies is limited. The results from the actual application of these techniques to classification of the moving objects of a road scene into groups and remaining anomalies is presented and analysed.