This project aims at developing new methodology for detecting features from receptive fields for handling time-dependent image data (video/real-time image streams) for purposes in basic motion analysis and for recognizing motion patterns in space-time. Specifically, we aim at developing scale-space theory for handling spatio-temporal image data (video and image streams) from real-time needs using time-causal image operations and compact time-recursive implementations.
Today, a major part of the method development is performed on pre-recorded video, where one can take the liberty of accessing the virtual future in relation to any pre-recorded time moment. Thereby, mostly non-causal temporal image operations are being used. For a genuine real-time system or for modelling biological vision, it is however required to base the analysis on truly time-causal video operations, which are only allowed to access data from the present moment and the past.
In this area, the theory developments have so far been rather limited, which implies a need for a focused research effort on time-causal and time-recursive receptive fields, which we aim at developing based on a recently developed generalized framework.
We aim at developing basic theory, build discrete implementations and evaluate these experimentally towards problems in motion analysis and spatio-temporal recognition (recognition of motion patterns). The results of the project are expected to be of a generic nature, with applicability to different problems in computer vision and related areas. The deeper theoretical understanding that the project aims at can also be expected to be applicable for understanding and modelling biological vision.