Pia Bideau’s X-Student Research Groups brings Lego robots and RaspberryPis to SCIoI
What do changes in brightness captured by an event camera tell us about distance? Last semester, the BUA-funded X-Student Research Group “Active Perception,” led by Pia Bideau, worked on finding the answer to this question. With this in mind, students further developed recent works by Pia Bideau and Guillermo Gallego and Aravind Battaje and Oliver Brock.
Current computer vision methods aim at processing an entire image at once. Prior work however has shown that the one-image-at-once approach might not be the most effective. Instead, biological vision explores a scene while consecutively focusing on different regions in the scene. Students in this X-Student Research Ggroup in particular explored alignment strategies for traditional rgb-videos and event streams recorded by an event camera, and how those could be beneficial for distance estimation.
In case of event data, event alignment has been typically used to align events over the entire image to get edge images that are much easier to interpret, and objects are much better to identify. As a by-product, event alignment can recover the camera’s motion. In this project course, students developed a novel alignment strategy that does not aim at aligning all events captured on the entire camera’s sensor, but instead only aligns events recorded within a small area of the camera’s sensor (comparable to aligning a small area of the image). While this does not allow for recovery of the camera’s motion, this strategy allows a very good estimate of distance. The developed approach to recover distance from motion was implemented and tested on pre-recorded event camera sequences as well as on a Legorobot equipped with an rgb-camera and RaspberryPi.
The X-Student Research Group “Active Perception” was supported by the Berlin-University Alliance. A follow-up course will be offered this coming summer semester.
Image: A scene captured by an event camera (image in middle) and the events aligned within a small image region for distance estimation (left image). The bright circle pictures the small area where events get aligned.
