Robust Multi-Scale Methods for Optic Flow (ENN.6760)

Research:
Medical applications, video applications in the consumer market, robot vision, traffic monitoring, industrial inspec-tion and various other applications of image processing technology require objects in video sequences to be seg-mented. In such applications typically a (typically small) number of objects have to be tracked in a continuously changing background environment.
There is an increasing need to extract these motion fields automatically. The amount of data becomes overwhelming, especially in the medical field, but the same is true for all fields mentioned above. Some medical examples:

• Extraction of the deformation field ('local warping') to register two images from different modalities or different moments, to integrate the complementary information, or to study changes (e.g. tumor growth) over time
• Quantitative analysis of ventricular heart wall motion, to study the mechanical properties of the infarcted heart, and to calculate the possible location of the infarcted region
• Motion vector fields also are exploited to arrive at a meaningful segmentation: a segment is then characterized by the fact that all its pixels have sufficiently uniform velocities

We propose in this proposal to investigate how scale space methods can be exploited to interrogate time dependent image streams. In particular, it is the intention to obtain stable, dense and accurate motion vector fields and to arrive at time-dependent segmentations based on such motion fields.

The utilization areas are chosen by the industrial partners, and are described in the next summary section.
The partners form an experienced and balanced team of experts in mathematics, multi-scale (biomedical) computer vision applications, video sequence analysis and X-ray fluoroscopic cardiovascular applications. Partners have col-laborated in a number of previous projects.

Utilisation
We distinguish short term and long-term utilisation.
In the long term, this research contributes to the representation of 2-D, 3-D, and 4-D signals (images, time se-quences of images, volume sets and time sequences of volume sets) at higher semantic levels. This means that mean-ingful, time-dependent features in the signals can be represented, transmitted, manipulated and stored explicitly. For example, identification of motion patterns on behalf of diagnostic purposes in medical applications; encoding of moving objects on behalf of efficient bandwidth utilisation in video, or motion detection in surveillance applications. This has large impact on a broad range of applications, ranging from entertainment and consumer applications (inter-active video, videophone, computer games, contents-based data retrieval for the Web, etc.) to professional applica-tions (pattern recognition and computer vision, simulation and augmented reality, etc.).

In the short term, we focus on two important utilization areas, each with a substantial market:

• Medical diagnosis, treatment planning, and intervention in moving tissue (e.g., diagnosis of heart disor-ders). The Philips group X-ray Development will utilize the developed techniques in the tracking of catheters, where the predicted motion path may lead to substantial dose reduction and improvement of the detec-tion of the catheter's precise location.
• Video encoding of moving segments ('video objects' as defined in MPEG4 context). Philips Research has a longstanding experience in dynamic video analysis. The developed techniques may lead to a highly efficient predictive coding of contextual motion fields, and may lead to efficient and robust, quantitative and dense, motion detection.