Towards Reconstructing Objects behind Fragmented Occlusion using Video

Description

Seeing through foliage is natural for the human visual system, but still challenging for vision algorithms. A recognition model that might also explain human capabilities is therefore needed, which will lead to many new applications. Reconstructing moving objects in the presence of fragmented occlusion is perhaps a first step towards a better understanding of such a model. An approach with synthetic apertures need 88 cameras for 70% occlusion density which makes reconstruction under even denser foliage difficult and is obviously not a plausible model of the human visual system. Using grey-scale images captured from a single camera, we investigated an alternative approach to this reconstruction problem. The paper suggests a new algorithm that registers the images with respect to the object, and then averages them to an integrated image which is a new representation of the occluded object. It explores the parameter space of the object's motion and estimates the optimal parameters maximising the contrast of the integrated image. Experiments with both synthetic data and real videos demonstrated the feasibility of reconstruction for occlusion densities even larger than 90%. As a consequence, motion should be taken into account in models of recognition. Still, the parametric motion, averaging and contrast maximisation are factors of limited application and will therefore remain the focus of future work.

Period	29 Nov 2024
Held at	NORCE Norwegian Research Centre, Norway
Degree of Recognition	International