Publicação
A bio-inspired computational model for motion detection
| Resumo: | Last years have witnessed a considerable interest in research dedicated to show that solutions to challenges in autonomous robot navigation can be found by taking inspiration from biology. Despite their small size and relatively simple nervous systems, insects have evolved vision systems able to perform the computations required for a safe navigation in dynamic and unstructured environments, by using simple, elegant and computationally efficient strategies. Thus, invertebrate neuroscience provides engineers with many neural circuit diagrams that can potentially be used to solve complicated engineering control problems. One major and yet unsolved problem encountered by visually guided robotic platforms is collision avoidance in complex, dynamic and inconstant light environments. In this dissertation, the main aim is to draw inspiration from recent and future findings on insect’s collision avoidance in dynamic environments and on visual strategies of light adaptation applied by diurnal insects, to develop a computationally efficient model for robotic control, able to work even in adverse light conditions. We first present a comparative analysis of three leading collision avoidance models based on a neural pathway responsible for signing collisions, the Lobula Giant Movement Detector/Desceding Contralateral Movement Detector (LGMD/DCMD), found in the locust visual system. Models are described, simulated and results are compared with biological data from literature. Due to the lack of information related to the way this collision detection neuron deals with dynamic environments, new visual stimuli were developed. Locusts Lo- custa Migratoria were stimulated with computer-generated discs that traveled along a combination of non-colliding and colliding trajectories, placed over a static and two distinct moving backgrounds, while simultaneously recording the DCMD activity extracellularly. Based on these results, an innovative model was developed. This model was tested in specially designed computer simulations, replicating the same visual conditions used for the biological recordings. The proposed model is shown to be sufficient to give rise to experimentally observed neural insect responses. Using a different approach, and based on recent findings, we present a direct approach to estimate potential collisions through a sequential computation of the image’s power spectra. This approach has been implemented in a real robotic platform, showing that distant dependent variations on image statistics are likely to be functional significant. Maintaining the collision detection performance at lower light levels is not a trivial task. Nevertheless, some insect visual systems have developed several strategies to help them to optimize visual performance over a wide range of light intensities. In this dissertation we address the neural adaptation mechanisms responsible to improve light capture on a day active insect, the bumblebee Bombus Terrestris. Behavioral analyses enabled us to investigate and infer about the spatial and temporal neural summation extent applied by those insects to improve image reliability at the different light levels. As future work, the collision avoidance model may be coupled with a bio-inspired light adaptation mechanism and used for robotic autonomous navigation. |
|---|---|
| Autores principais: | Silva, Ana Carolina Quintela Alves Vilares da |
| Assunto: | Engenharia e Tecnologia::Biotecnologia Industrial |
| Ano: | 2015 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Last years have witnessed a considerable interest in research dedicated to show that solutions to challenges in autonomous robot navigation can be found by taking inspiration from biology. Despite their small size and relatively simple nervous systems, insects have evolved vision systems able to perform the computations required for a safe navigation in dynamic and unstructured environments, by using simple, elegant and computationally efficient strategies. Thus, invertebrate neuroscience provides engineers with many neural circuit diagrams that can potentially be used to solve complicated engineering control problems. One major and yet unsolved problem encountered by visually guided robotic platforms is collision avoidance in complex, dynamic and inconstant light environments. In this dissertation, the main aim is to draw inspiration from recent and future findings on insect’s collision avoidance in dynamic environments and on visual strategies of light adaptation applied by diurnal insects, to develop a computationally efficient model for robotic control, able to work even in adverse light conditions. We first present a comparative analysis of three leading collision avoidance models based on a neural pathway responsible for signing collisions, the Lobula Giant Movement Detector/Desceding Contralateral Movement Detector (LGMD/DCMD), found in the locust visual system. Models are described, simulated and results are compared with biological data from literature. Due to the lack of information related to the way this collision detection neuron deals with dynamic environments, new visual stimuli were developed. Locusts Lo- custa Migratoria were stimulated with computer-generated discs that traveled along a combination of non-colliding and colliding trajectories, placed over a static and two distinct moving backgrounds, while simultaneously recording the DCMD activity extracellularly. Based on these results, an innovative model was developed. This model was tested in specially designed computer simulations, replicating the same visual conditions used for the biological recordings. The proposed model is shown to be sufficient to give rise to experimentally observed neural insect responses. Using a different approach, and based on recent findings, we present a direct approach to estimate potential collisions through a sequential computation of the image’s power spectra. This approach has been implemented in a real robotic platform, showing that distant dependent variations on image statistics are likely to be functional significant. Maintaining the collision detection performance at lower light levels is not a trivial task. Nevertheless, some insect visual systems have developed several strategies to help them to optimize visual performance over a wide range of light intensities. In this dissertation we address the neural adaptation mechanisms responsible to improve light capture on a day active insect, the bumblebee Bombus Terrestris. Behavioral analyses enabled us to investigate and infer about the spatial and temporal neural summation extent applied by those insects to improve image reliability at the different light levels. As future work, the collision avoidance model may be coupled with a bio-inspired light adaptation mechanism and used for robotic autonomous navigation. |
|---|