This research focuses on dynamic high-speed vision based on temporal differential operators for dynamic scene analysis and motion estimation. Dynamic vision for autonomous systems exploiting the visual flow field is of particular advantage in weakly structured environments with lack of classifiable objects. The challenge lies in the extraction of the flow field inducing motion of the cognitive system, respectively, of moving objects and object components in the scene. Stochastic observers are a common approach for ego-motion estimation, however, computationally expensive due to extensive matrix multiplications.
A number of flying insects, e.g. Drosophila melanogaster and Calliphora vicina, possess photoreceptors with high temporal resolution which they use for dynamic visuomotor pose and gaze stabilization, and navigation in 6 degrees of freedom. Our research addresses novel aspects in neurobiological foundations of Calliphora's perceptual processes and visuomotor behaviors, and the technological realization and integration with methods from control engineering and computer vision. It benefits from synergies between fundamental research and technology.
In the first phase, fundamentals of vision-based motion estimation of Calliphora as well as transfer and implementation of neurobiological results are envisaged. Technical approaches in high-speed image processing, vision-based control, multi-rate sensor data fusion, their integration with the computational models from neurobiology, and implementations on selected demonstrators as the Quadrotor complete this phase. The contribution is the enhancement of perceptual capabilities of cognitive systems by perceiving dynamic visual information for dynamic map building and evaluation of object motions. The medium-term perspective is the extraction of scene motion information by subtracting ego-motion contributing to cognitive systems not only in a mapping, but also especially in a recognition context in a framework , e.g. analyzing human behaviors, emotions, and intentions evaluating human body motions.
Researcher
Kolja Kühnlenz
Tianguang Zhang