Features for matching people in different views

Abstract

There have been significant advances in the computer vision field during the last decade. During this period, many methods have been developed that have been successful in solving challenging problems including Face Detection, Object Recognition and 3D Scene Reconstruction. The solutions developed by computer vision researchers have been widely adopted and used in many real-life applications such as those faced in the medical and security industry. Among the different branches of computer vision, Object Recognition has been an area that has advanced rapidly in recent years. The successful introduction of approaches such as feature extraction and description has been an important factor in the growth of this area. In recent years, researchers have attempted to use these approaches and apply them to other problems such as Content Based Image Retrieval and Tracking. In this work, we present a novel system that finds correspondences between people seen in different images. Unlike other approaches that rely on a video stream to track the movement of people between images, here we present a feature-based approach where we locate a target’s new location in an image, based only on its visual appearance. Our proposed system comprises three steps. In the first step, a set of features is extracted from the target’s appearance. A novel algorithm is developed that allows extraction of features from a target that is particularly suitable to the modelling task. In the second step, each feature is characterised using a combined colour and texture descriptor. Inclusion of information relating to both colour and texture of a feature add to the descriptor’s distinctiveness. Finally, the target’s appearance and pose is modelled as a collection of such features and descriptors. This collection is then used as a template that allows us to search for a similar combination of features in other images that correspond to the target’s new location. We have demonstrated the effectiveness of our system in locating a target’s new position in an image, despite differences in viewpoint, scale or elapsed time between the images. The characterisation of a target as a collection of features also allows our system to robustly deal with the partial occlusion of the target.