Uploaded April 2019 | Updated November 2026, 3 hours ago
Paper: ieeexplore.ieee.org/abstract/document/8660483
arxiv.org/abs/1807.10936
cuSNN library: github.com/tudelft/cuSNN
We present the first SNN in which selectivity to the local and global motion of the visual scene emerges through STDP from event-based stimuli. The success of this emergence depends on three contributions. First, an adaptive spiking neuron model is necessary to handle the rapidly varying input statistics of event-based sensors, and we present a novel suitable formulation for this purpose. Second, we introduce a novel STDP implementation that, contrary to the current state-of-the-art of this learning protocol, is inherently stable. Third, we propose an SNN architecture that learns to perform hierarchical feature extraction, effectively capturing geometric features, identifying the local motion of these features, and integrating this information into a global ego-motion estimate. We hope that this work, and the framework published alongside it, will provide the first step of many towards highly efficient artificial motion perception.
Paper: ieeexplore.ieee.org/abstract/document/8660483
arxiv.org/abs/1807.10936
cuSNN library: github.com/tudelft/cuSNN
We present the first SNN in which selectivity to the local and global motion of the visual scene emerges through STDP from event-based stimuli. The success of this emergence depends on three contributions. First, an adaptive spiking neuron model is necessary to handle the rapidly varying input statistics of event-based sensors, and we present a novel suitable formulation for this purpose. Second, we introduce a novel STDP implementation that, contrary to the current state-of-the-art of this learning protocol, is inherently stable. Third, we propose an SNN architecture that learns to perform hierarchical feature extraction, effectively capturing geometric features, identifying the local motion of these features, and integrating this information into a global ego-motion estimate. We hope that this work, and the framework published alongside it, will provide the first step of many towards highly efficient artificial motion perception.








![Tudelft bebop paparazzi
As you know paparazzi is very modular platform. To show the world how modular we are we implementing paparazzi into the Parrot Bebop. At the Micro Aerial Vehicle Laboratory ([http://www.http://mavlab.lr.tudelft.nl/]) of Delft University of Technology faculty Aerospace engineering ([http://http://www.lr.tudelft.nl/ TU Delft]) we received a pre-production model of the Bebop. We do best to have paparazzi integrated before the launch of the Bebop. If you want to follow the progress check our youtube channel Tudelft bebop paparazzi](https://i.ytimg.com/vi/HEa9lcDIO2g/mqdefault.jpg)

