Spatio-temporal volume-based shape modelling for video event detection

In a typical computer vision application, such as video event detection, the 'meaningful' information is fundamentally represented by pre-defined features, which determine the appropriate analytical methodologies in the following processing phases. Based on the uncompressed low-level image characteristics, such as colour, intensity and spatial positions, the features used for event detection in this research are predominantly based on 3D shapes, regional textures, and sudden colour/intensity. In this research, a spatio-temporal volume-based shape feature extraction and modelling approach has been proposed. This method starts from defining video data as 3D volumetric shapes by using active contour (AC) segmentation techniques. Based on the nature of its 3D distribution, a dynamic windowing mechanism has been developed for improving the segmentation performance when deploying the AC algorithm. The runtime performance of the prototype system has been evaluated which validated the design and its potential in improving volume-based event recognition.