Neuromorphic Vision Datasets for Pedestrian Detection, Action Recognition, and Fall Detection

• Published in: Frontiers in neurorobotics 2019-06, Vol.13, p.38-38
• Main Authors: Miao, Shu, Chen, Guang, Ning, Xiangyu, Zi, Yang, Ren, Kejia, Bing, Zhenshan, Knoll, Alois
• Format: Article
• Language: English
• Subjects: action recognition; Cameras; dataset; Datasets; dynamic vision sensor; fall detection; Neural networks; Neuroscience; Pattern recognition; pedestrian detection; Robots; Sensors; Vision systems
• ISSN: 1662-5218; 1662-5218
• DOI: 10.3389/fnbot.2019.00038

Many of them were recorded with a static DVS facing a monitor on which computer vision datasets were set to play automatically (Serrano-Gotarredona and Linares-Barranco, 2015; Hu et al., 2016). [...]the intrinsic temporal information of moving objects between two frames are lost. [...]we aim to fill this gap and to introduce three datasets in this report: the pedestrian detection dataset, the action recognition dataset and the fall detection dataset. Surface of Active Events (SAE) In order to take full advantage of the unique characteristic that neuromorphic vision sensors can record the exact occurring time of incoming events with low latency, the SAE (Surface of Active Events) (Mueggler et al., 2017b) approach is applied to reflect time information while the pixel value and its gradient can tell the moving direction and speed of the event stream. Specifically, regardless of the event polarity, each incoming event [t, x, y, p] will change the pixel value tp at (x, y) according to the timestamp t. In this way, an grayscale image frame is acquired according to the timestamp of the most recent event at each pixel: SAE:t⇒tp(x,y) (2) Moreover, to attain an 8-bit single channel image, numerical mapping is conducted by calculating the Δt between the pixel value tp and the initial time t0 for each frame interval T as follows: g(x,y)=255·tp-t0T (3) 2.3.3.